https://radar.trulysovereign.ai RADAR items in the AI Platform Drift channel. en Thu, 24 Sep 2026 04:30:30 GMT https://radar.trulysovereign.ai/items/0188db92-85d1-4c58-bbc3-c02bbe9d9041 0188db92-85d1-4c58-bbc3-c02bbe9d9041 Thu, 24 Sep 2026 03:31:36 GMT budget OpenRouter model catalogue OpenRouter's deepseek-pro-latest model changed pricing: prompt tokens decreased slightly from $0.00000038808 to $0.0000003865 per token, but completion tokens increased from $0.00000116424 to $0.0000029 per token (149% increase). Maximum completion tokens increased from 393,216 to 943,718 tokens. Why it matters: If you generate long outputs with this model, your per-request costs will more than double. A 100k token completion that previously cost $0.116 now costs $0.29. The higher output limit means requests that previously hit the cap can now run longer and cost proportionally more. What to do: Calculate your typical completion token usage for this model and reforecast your API budget. If you rely on the output limit as a cost control, set explicit max_tokens parameters in your requests to prevent unexpectedly large bills. https://radar.trulysovereign.ai/items/156b2a0d-978e-48ae-ac5e-e5c9f4a169af 156b2a0d-978e-48ae-ac5e-e5c9f4a169af Thu, 24 Sep 2026 03:31:21 GMT verify_only OpenRouter model catalogue Effective now: prompt pricing decreased from $0.15 to $0.14 per million tokens, completion pricing decreased from $0.60 to $0.42 per million tokens (30% reduction). Maximum completion tokens reduced from 393,216 to 131,072. Why it matters: If you generate responses longer than 131,072 tokens, calls will now fail or truncate. The pricing drop reduces costs for existing workloads but the token limit is a breaking change for long-form generation. What to do: Check your application logs for any DeepSeek V4.1 Flash responses exceeding 131,072 tokens in the past 30 days. If found, either switch to a model with higher limits or redesign the prompt to stay under the new ceiling. https://radar.trulysovereign.ai/items/2d9fae4e-86a9-4c20-ab71-7e46259098e5 2d9fae4e-86a9-4c20-ab71-7e46259098e5 Wed, 23 Sep 2026 21:13:08 GMT budget OpenRouter model catalogue Prompt pricing decreased from $0.0000004 to $0.00000038808 per token (3% reduction). Completion pricing decreased from $0.0000043 to $0.00000116424 per token (73% reduction). Maximum completion tokens increased from 384,000 to 393,216 tokens. Why it matters: If you are using this model, your completion costs just dropped by nearly three-quarters. For workloads generating long outputs, the combination of lower per-token cost and higher output ceiling means significantly lower cost per request and fewer truncation issues. What to do: Pull your last 30 days of usage for deepseek-pro-latest from OpenRouter logs. Multiply completion token volume by the old and new rates to quantify the savings, then update your cost forecast and notify whoever owns the AI budget. https://radar.trulysovereign.ai/items/1eb8d1ec-3a98-44e2-adf4-bc6ffb7c8ce4 1eb8d1ec-3a98-44e2-adf4-bc6ffb7c8ce4 Wed, 23 Sep 2026 21:12:46 GMT verify_only OpenRouter model catalogue The context_length for inclusionAI: Ling 3.0 Flash VL increased from 131072 to 262144 tokens. Pricing remains unchanged at $0.00000006 per prompt token and $0.00000018 per completion token. Why it matters: If you currently truncate or chunk inputs to fit the old 131k limit, you can now send twice as much context in a single call, reducing multi-turn overhead and improving coherence for long-document tasks. What to do: Grep your codebase for '131072' or references to this model's context limit and evaluate whether batching logic or chunking strategies can be simplified to use the full 262k window. https://radar.trulysovereign.ai/items/f4367632-9ebd-4f1d-8da7-c21f1c617f7b f4367632-9ebd-4f1d-8da7-c21f1c617f7b Wed, 23 Sep 2026 21:12:37 GMT budget OpenRouter model catalogue OpenRouter's DeepSeek V4.1 Flash model pricing increased from $0.0000001 to $0.00000015 per prompt token (50% increase) and $0.0000005 to $0.0000006 per completion token (20% increase). Maximum completion tokens dropped from 943,718 to 393,216 tokens (58% reduction). Why it matters: Applications generating long outputs will hit the new 393K token ceiling where they previously could produce up to 943K tokens. Monthly costs will increase proportionally for all usage: a workload consuming 1B prompt tokens and 200M completion tokens now costs $270/month instead of $200/month. What to do: Calculate your current monthly token usage for this model and reforecast costs with the new rates. Check whether any workflows rely on outputs exceeding 393,216 tokens and either chunk the requests or switch models. https://radar.trulysovereign.ai/items/f654bd7b-a520-4e76-973f-111f69f761f2 f654bd7b-a520-4e76-973f-111f69f761f2 Wed, 23 Sep 2026 21:12:27 GMT verify_only Azure Foundry model retirement schedule The retirement date of 2026-08-31 was removed from gpt-realtime-2 (preview model launched 2026-05-06). A new model version gpt-realtime-2.1 was added to the schedule with a launch date of 2026-07-07 and no stated retirement date. Why it matters: If you are using gpt-realtime-2 and planned migration work around the August 31 deadline, that deadline no longer exists. The model remains in preview with no published end-of-life, which means the retirement timeline is now undefined. What to do: Check your deployment configs for gpt-realtime-2 references. If you scheduled migration work for before August 31, deprioritize it until a new retirement date is published or evaluate whether to migrate to gpt-realtime-2.1 based on your release criteria for preview models. https://radar.trulysovereign.ai/items/9d3c2cca-9339-4d39-aaee-5059fb7eb195 9d3c2cca-9339-4d39-aaee-5059fb7eb195 Wed, 23 Sep 2026 21:12:12 GMT migrate Claude Platform release notes Claude Opus 5.5 (claude-opus-5-5) launched with 1M token context, 128k max output, $4/$20 per MTok (down from Opus 5's $5/$25). On this model, thinking: {"type": "disabled"} and thinking: {"type": "enabled"} now return 400 errors; tool_choice types "any" and "tool" also return 400. On Claude API and Google Cloud, computer_20251124 tool returns 400 and requires computer_toolset_20260801 instead. Amazon Bedrock still accepts computer_20251124. Why it matters: Existing code that sets thinking parameters or uses tool_choice "any"/"tool" will break with 400 errors if you switch to Opus 5.5. Code using computer_20251124 on Claude API or Google Cloud will also break. If you have hardcoded model names or auto-upgrade logic, you may hit these errors without intending to migrate. What to do: Grep your codebase for "thinking": {"type" and tool_choice references. If found, pin to claude-opus-5 in model selection logic until you can remove those parameters and migrate computer use toolsets per the migration guide. If you do not use these features, verify your model selection does not auto-upgrade, then test Opus 5.5 in staging to confirm 20% cost savings apply to your workload. https://radar.trulysovereign.ai/items/61873d88-7667-4f7e-9bb7-d177ca8c8fe0 61873d88-7667-4f7e-9bb7-d177ca8c8fe0 Wed, 23 Sep 2026 16:16:55 GMT budget OpenRouter model catalogue Z.ai GLM 5.3 batch model pricing decreased from $0.00000072 to $0.00000045 per prompt token (37.5% reduction) and from $0.0000024 to $0.000002 per completion token (16.7% reduction), effective as of 23 September 2026. Why it matters: If you are using this model in batch mode, your per-request costs have dropped by roughly one-third on the prompt side. Existing cost forecasts and budget allocations for this model are now overstated. What to do: Rerun cost projections for any workloads using z-ai/glm-5.3:batch and update budget forecasts to reflect the new rates. Notify whoever owns the AI spend line item. https://radar.trulysovereign.ai/items/6c0172e7-7c08-4c28-963d-a2074a2356f7 6c0172e7-7c08-4c28-963d-a2074a2356f7 Wed, 23 Sep 2026 10:56:49 GMT budget OpenRouter model catalogue OpenRouter raised DeepSeek V4 Flash Latest pricing from $0.00000003 to $0.000000038 per prompt token (+27%) and from $0.000001 to $0.00000055 per completion token (-45%). Change is live as of 23 September 2026. Why it matters: If you are using this model in production, your inference costs decreased by approximately 45% for completion-heavy workloads but increased 27% for prompt-heavy workloads. Net impact depends on your prompt-to-completion token ratio. What to do: Pull last 30 days of usage logs for this model, calculate actual prompt vs completion token distribution, and reforecast monthly spend using the new rates. If completion tokens dominate (typical for generation tasks), you saved money; if prompts dominate (classification, embeddings), costs went up. https://radar.trulysovereign.ai/items/97e7f9c9-337b-4a0b-9b82-aac3000973da 97e7f9c9-337b-4a0b-9b82-aac3000973da Wed, 23 Sep 2026 10:56:17 GMT verify_only OpenRouter model catalogue Prompt pricing increased from $0.00000007 to $0.00000008 per token (14% increase). Maximum completion tokens reduced from 235,929 to 131,072 tokens (44% reduction). Changes are live as of 23 September 2026. Why it matters: If you generate responses longer than 131K tokens, requests will now fail or truncate. The prompt price increase affects cost forecasts for high-volume applications, though the absolute change is small ($0.01 per 100K tokens). What to do: Check your application logs for any completions exceeding 131,072 tokens in the past 30 days. If found, either chunk your outputs or switch models. Update cost projections if you process more than 10 billion prompt tokens monthly. https://radar.trulysovereign.ai/items/ad095edd-df01-4fac-a278-0a84ed315743 ad095edd-df01-4fac-a278-0a84ed315743 Wed, 23 Sep 2026 03:42:44 GMT verify_only OpenRouter model catalogue The top_provider.max_completion_tokens field dropped from 943,718 to 131,072 tokens—an 86% reduction. Context length remains 1,310,720 tokens. Pricing unchanged at $0.000000075 per prompt token and $0.00000025 per completion token. Why it matters: Any request attempting to generate more than 131,072 tokens will now fail or be truncated. If you rely on long-form generation (summarizing large documents, generating extensive code, multi-turn conversations with large outputs), existing calls may break without warning. What to do: Grep your codebase for max_tokens or max_completion_tokens parameters sent to OpenRouter. If any exceed 131,072, reduce them or split the work into multiple requests. Test one long-output workflow end-to-end this week. https://radar.trulysovereign.ai/items/9895f5ac-8a22-4a5d-aa65-b07795edd647 9895f5ac-8a22-4a5d-aa65-b07795edd647 Wed, 23 Sep 2026 03:42:35 GMT budget OpenRouter model catalogue Effective now, DeepSeek Pro Latest completion pricing increased from $0.0000012/token to $0.0000043/token (3.6x). Prompt pricing rose marginally from $0.0000004 to $0.0000004. Maximum completion tokens dropped from 393,216 to 384,000 (9,216 tokens less). Why it matters: Any application generating long outputs will see immediate cost increases—a 100k token completion now costs $0.43 instead of $0.12. Workloads relying on the full 393k output window will fail or truncate. What to do: Pull last 30 days of DeepSeek Pro usage from OpenRouter, calculate new cost at $0.0000043/completion token, and update your budget forecast. If any jobs request >384k output tokens, either split them or switch models before they hit the new limit. https://radar.trulysovereign.ai/items/c69827bd-de49-4ccd-a64d-1a1b7a7da9ea c69827bd-de49-4ccd-a64d-1a1b7a7da9ea Wed, 23 Sep 2026 03:42:25 GMT budget OpenRouter model catalogue Effective now: prompt pricing decreased from $0.15/M to $0.10/M tokens (33% reduction), completion pricing decreased from $0.60/M to $0.50/M tokens (17% reduction). Maximum completion tokens increased from 384,000 to 943,718. Why it matters: If you are using this model, your per-request costs just decreased by roughly 20-30% depending on prompt/completion ratio. If you previously hit the 384K completion limit, you can now generate responses 2.5x longer without chunking. What to do: Pull last 30 days of usage from your OpenRouter dashboard, recalculate total cost at new rates, and update your budget forecast. If you avoided this model due to output length constraints, re-evaluate it for long-form generation tasks. https://radar.trulysovereign.ai/items/00e9446c-9533-47e7-b05f-055b8713668b 00e9446c-9533-47e7-b05f-055b8713668b Tue, 22 Sep 2026 20:59:45 GMT budget OpenRouter model catalogue OpenRouter reduced pricing for deepseek/deepseek-pro-latest from $0.00000056364 to $0.00000039996 per prompt token (29% decrease) and $0.00000169092 to $0.00000119988 per completion token (29% decrease). Maximum completion tokens increased from 384,000 to 393,216. Why it matters: If you are using this model, your API costs will drop by roughly 29% with no code changes required. The higher output limit means requests that previously hit the 384K token ceiling can now complete without truncation. What to do: Reforecast your OpenRouter spend for this model downward by 29% and notify whoever owns the budget line. If you have retry logic or truncation handling for 384K token limits, verify it still works correctly at the new 393K ceiling. https://radar.trulysovereign.ai/items/8894ab07-4e7e-45d1-a4e1-3169d3e2c63b 8894ab07-4e7e-45d1-a4e1-3169d3e2c63b Tue, 22 Sep 2026 20:59:33 GMT budget OpenRouter model catalogue OpenRouter dropped pricing for anthropic/claude-opus-latest from $0.000005 to $0.000004 per prompt token (20% reduction) and from $0.000025 to $0.00002 per completion token (20% reduction), effective now. Why it matters: Any application currently using this model will see immediate cost reduction. Existing budget forecasts and cost alerts based on the old rates are now overstated by 20%. What to do: Update cost projections and budget models to reflect the new rates. If you have spend alerts or cost dashboards for this model, adjust thresholds down by 20% to maintain the same real-dollar triggers. https://radar.trulysovereign.ai/items/8293e819-5a08-45f0-b956-7d0300c6f0d6 8293e819-5a08-45f0-b956-7d0300c6f0d6 Tue, 22 Sep 2026 20:59:12 GMT verify_only OpenRouter model catalogue OpenRouter added openai/gpt-6-sol:batch with $0.000001 per prompt token, $0.000005 per completion token, 1.05M token context window, and 128k completion token limit. Why it matters: This is a new batch processing option for GPT-6 Sol at significantly lower pricing than real-time inference, enabling cost-effective processing of large-scale or non-time-sensitive workloads through OpenRouter's unified API. What to do: If you process GPT-6 Sol requests that can tolerate batch latency, test this endpoint with a sample workload to measure cost savings against your current real-time usage. https://radar.trulysovereign.ai/items/696c786e-6f91-4444-89f2-2ae9c172591f 696c786e-6f91-4444-89f2-2ae9c172591f Tue, 22 Sep 2026 20:59:04 GMT verify_only OpenRouter model catalogue OpenRouter added openai/gpt-6-sol-pro:batch to their catalogue on 2026-09-22. Pricing is $0.000001 per prompt token and $0.000005 per completion token. Context length is 1,050,000 tokens with a 128,000 token completion limit. Why it matters: Batch processing typically offers 50% cost savings over synchronous API calls for non-time-sensitive workloads. The 1M+ context window enables processing entire codebases or long documents in a single request without chunking. What to do: If you currently use gpt-4 or gpt-4-turbo for batch jobs, benchmark gpt-6-sol-pro:batch against your workload and calculate cost difference at the stated rates. Update your model routing logic to include this identifier if the quality-cost tradeoff works. https://radar.trulysovereign.ai/items/24322636-3796-4d2c-90be-dc690db435d6 24322636-3796-4d2c-90be-dc690db435d6 Tue, 22 Sep 2026 20:58:55 GMT verify_only OpenRouter model catalogue OpenRouter added openai/gpt-6-sol-pro to their catalogue on 2026-09-22. Pricing is $0.000002 per prompt token and $0.00001 per completion token. Context length is 1,050,000 tokens with a 128,000 token completion limit. Why it matters: This is a new model tier with significantly expanded context (1M+ tokens) at rates 5x higher for completions than typical GPT-4 pricing. If your application would benefit from the larger context window, you now have a routing option; if cost optimization matters more, this is not the model to default to. What to do: Check whether any of your OpenRouter integrations use dynamic model selection or 'auto' routing that might pick up this model. If so, verify the cost implications before it gets selected in production traffic. https://radar.trulysovereign.ai/items/7249cb54-0cb6-45fc-ba04-8f99ba10bf9c 7249cb54-0cb6-45fc-ba04-8f99ba10bf9c Tue, 22 Sep 2026 20:58:41 GMT verify_only OpenRouter model catalogue OpenRouter added openai/gpt-6-luna:batch with pricing at $0.00000005 per prompt token and $0.00000025 per completion token, 1,050,000 token context window, and 128,000 token completion limit. Why it matters: This batch endpoint offers significantly lower pricing than typical real-time models for workloads that can tolerate asynchronous processing, and the 1M+ context window enables processing very large documents or conversations in a single call. What to do: If you have batch processing workloads or need to process documents larger than your current context limits, test this model against your existing pipeline to measure cost savings and latency trade-offs. https://radar.trulysovereign.ai/items/5de430ed-bf81-422c-8c6a-641b3d74c38c 5de430ed-bf81-422c-8c6a-641b3d74c38c Tue, 22 Sep 2026 20:58:13 GMT budget OpenRouter model catalogue OpenRouter added anthropic/claude-opus-5.5:batch with prompt tokens at $0.000002 and completion at $0.00001, 1M context window, 128k max completion tokens. This is half the cost of the standard Opus 5.5 model. Why it matters: If you run high-volume or non-latency-sensitive Claude Opus 5.5 workloads, batch mode cuts your API costs in half. Batch processing typically adds delay but reduces per-token cost by accepting asynchronous execution. What to do: Identify any Claude Opus 5.5 calls that can tolerate async processing (report generation, bulk analysis, overnight jobs) and test switching those to the :batch endpoint to halve costs on that traffic. https://radar.trulysovereign.ai/items/a14958a8-b11b-4c2c-a9e2-86f70cdecc21 a14958a8-b11b-4c2c-a9e2-86f70cdecc21 Tue, 22 Sep 2026 16:28:36 GMT verify_only OpenRouter model catalogue The top_provider.max_completion_tokens for model ~z-ai/glm-latest dropped from 943,718 to 131,072 tokens (86% reduction). Pricing changed negligibly: prompt tokens from $0.0000006545 to $0.0000006538, completion tokens from $0.000002057 to $0.0000020548 per token. Why it matters: Any code requesting completion lengths above 131,072 tokens will now fail or be truncated. If you generate long-form content, transcripts, or batch outputs with this model, those requests will break. What to do: Grep your codebase for max_tokens or max_completion_tokens parameters sent to ~z-ai/glm-latest. If any exceed 131,072, either reduce them or switch to a model with higher limits before your next deployment. https://radar.trulysovereign.ai/items/e0c7de9b-2257-485c-ac70-e6c7e9cec208 e0c7de9b-2257-485c-ac70-e6c7e9cec208 Tue, 22 Sep 2026 16:28:27 GMT verify_only OpenRouter model catalogue Prompt pricing dropped from $0.000000624 to $0.00000056364 per token (~10% decrease). Completion pricing dropped from $0.00000288 to $0.00000169092 per token (~41% decrease). Maximum completion tokens reduced from 943,718 to 384,000 tokens (59% reduction). Changes are live as of 22 September 2026. Why it matters: If you generate responses longer than 384,000 tokens, requests will now fail or truncate where they previously succeeded. The cost reduction is real but the capability ceiling dropped by more than half. What to do: Check your application logs for any DeepSeek Pro responses exceeding 384,000 tokens in the past 30 days. If found, either switch to a model with higher limits or redesign the workflow to chunk outputs below the new ceiling. https://radar.trulysovereign.ai/items/f03f2087-0018-42ea-940a-06772654a4a4 f03f2087-0018-42ea-940a-06772654a4a4 Tue, 22 Sep 2026 16:28:18 GMT migrate OpenRouter model catalogue The Z.ai GLM 5.3 batch model reduced max_completion_tokens from 943,718 to 131,072 tokens (86% reduction). Prompt pricing increased from $0.0000007 to $0.00000072 per token (2.9% increase). Completion pricing increased from $0.0000022 to $0.0000024 per token (9.1% increase). Context length remains 1,048,576 tokens. Why it matters: Any batch job generating outputs longer than 131K tokens will now fail or truncate. If you were relying on the previous 943K output capacity for long-form generation, summarization of large documents, or multi-document synthesis, those workflows will break. The pricing increase affects cost forecasts for high-volume batch processing. What to do: Grep your codebase for 'glm-5.3:batch' and check max_tokens or completion length assumptions. If any job expects outputs over 131K tokens, either chunk the work into multiple calls or switch models. Recalculate batch processing costs using the new rates and update budgets if this model handles significant volume. https://radar.trulysovereign.ai/items/4170e7a8-8e1a-4479-ac56-b34fadd5942a 4170e7a8-8e1a-4479-ac56-b34fadd5942a Tue, 22 Sep 2026 16:28:05 GMT verify_only OpenRouter model catalogue OpenRouter's Z.ai GLM 5.3 Flash batch model reduced prompt pricing from $0.000000075 to $0.00000006 per token (20% decrease) and completion pricing from $0.00000025 to $0.0000002 per token (20% decrease). Simultaneously, max_completion_tokens dropped from 943,718 to 131,072 tokens (86% reduction). Why it matters: Any workload generating responses longer than 131,072 tokens will now fail or truncate. The pricing reduction does not compensate if you must switch to a different model to preserve output length capability. What to do: Query your logs for calls to z-ai/glm-5.3-flash:batch with completion lengths exceeding 131,072 tokens. If found, test whether your use case tolerates the new limit or requires migration to a model with higher max_completion_tokens before your next production run. https://radar.trulysovereign.ai/items/0737b2da-084d-4ce8-959e-72d583f9f303 0737b2da-084d-4ce8-959e-72d583f9f303 Tue, 22 Sep 2026 16:27:54 GMT budget OpenRouter model catalogue Effective now: prompt pricing increased from $0.0000003 to $0.00000032 per token (~6.7% increase), completion pricing increased from $0.000002 to $0.0000027 per token (35% increase). Maximum completion tokens increased from 65,536 to 262,140. Why it matters: If you are using this model in production, your per-request costs just increased by 35% on the output side. High-volume applications will see material budget impact. The 4x increase in max completion tokens enables longer responses but does not offset the price increase. What to do: Pull usage logs for qwen/qwen3.6-27b from the past 30 days, calculate new monthly cost at the updated rates, and update your budget forecast. If cost is material, evaluate whether switching to a different model in the Qwen family or another provider offers better economics. https://radar.trulysovereign.ai/items/e75db6f3-25c9-431d-a1d2-9f4474f44fc6 e75db6f3-25c9-431d-a1d2-9f4474f44fc6 Tue, 22 Sep 2026 16:27:40 GMT verify_only OpenRouter model catalogue OpenRouter's gpt-oss-20b model pricing dropped from $0.00000003 to $0.00000018 per prompt token (40% reduction) and $0.00000013 to $0.00000009 per completion token (31% reduction). Maximum completion tokens reduced from 117,964 to 32,768, effective now. Why it matters: If you generate responses longer than 32K tokens, calls will now truncate or fail. Cost forecasts based on the old pricing are overstated by roughly a third. What to do: Check your application logs for any gpt-oss-20b completions exceeding 32,768 tokens in the past 30 days. If found, either chunk your outputs or switch models. Update your cost model with the new per-token rates. https://radar.trulysovereign.ai/items/56fe4556-5c07-44fd-ad60-0b4cc4ae5732 56fe4556-5c07-44fd-ad60-0b4cc4ae5732 Tue, 22 Sep 2026 16:27:18 GMT verify_only OpenRouter model catalogue The context_length for AionLabs: Aion-3.0-Mini increased from 131,072 tokens to 1,048,576 tokens (8x expansion). Pricing remains $0.0000007 per prompt token and $0.0000014 per completion token. Why it matters: Applications currently chunking or truncating inputs to fit the 131K limit can now process 8x more content in a single request, eliminating multi-pass workflows and reducing latency. This changes the cost-performance calculation for document analysis, long-context retrieval, and conversation history retention. What to do: Identify pipelines currently splitting inputs for this model and benchmark end-to-end cost and latency with single-pass processing at the new limit. Update any hardcoded context_length checks from 131072 to 1048576. https://radar.trulysovereign.ai/items/f95482df-2648-4a33-8d2c-abe8e6384c60 f95482df-2648-4a33-8d2c-abe8e6384c60 Tue, 22 Sep 2026 16:27:08 GMT verify_only OpenRouter model catalogue The context_length for AionLabs: Aion-3.0 increased from 131,072 tokens to 1,048,576 tokens (8x expansion). Pricing remains unchanged at $0.000003 per prompt token and $0.000006 per completion token. Why it matters: Applications currently chunking or summarizing documents to fit within 131K tokens can now process 8x larger inputs in a single call, reducing complexity and potentially lowering total costs by eliminating multi-pass operations. What to do: Identify code paths that chunk documents or conversations specifically for Aion-3.0's previous 131K limit, and test whether removing that chunking logic improves performance or reduces API calls. https://radar.trulysovereign.ai/items/5c4c17cc-994a-45a1-b88b-4765bdc61b2f 5c4c17cc-994a-45a1-b88b-4765bdc61b2f Tue, 22 Sep 2026 16:27:00 GMT verify_only OpenRouter model catalogue The context_length for AionLabs: Aion-2.0 increased from 131,072 tokens to 1,048,576 tokens (8x expansion). Pricing remains unchanged at $0.0000008 per prompt token and $0.0000016 per completion token. Why it matters: Applications currently chunking or truncating inputs to fit the 131K limit can now process 8x more context in a single call, potentially eliminating multi-pass strategies and reducing total API calls for large document processing. What to do: Identify codepaths that chunk inputs for this model and evaluate whether removing chunking logic would reduce latency or cost. Test with representative large inputs to confirm the expanded window works as expected. https://radar.trulysovereign.ai/items/e0d67345-7065-4932-a736-dd92b09f4f68 e0d67345-7065-4932-a736-dd92b09f4f68 Tue, 22 Sep 2026 11:06:46 GMT verify_only OpenRouter model catalogue The top_provider.max_completion_tokens limit increased from 102,400 to 943,718 tokens (9.2x increase). Context length remains 1,310,720 tokens. Pricing unchanged at $0.000000075 per prompt token and $0.00000025 per completion token. Why it matters: If you currently cap completion requests at the old 102,400 limit in your code, you are leaving 841,318 tokens of available capacity unused. Conversely, if you request completions without explicit limits, responses could now be 9x larger than before, increasing latency and cost per request by the same factor. What to do: Grep your codebase for max_tokens or max_completion_tokens parameters in OpenRouter calls. If you set explicit caps, decide whether to raise them. If you rely on defaults, add explicit caps now to prevent unexpectedly large responses. https://radar.trulysovereign.ai/items/28edb445-2abd-4cba-9143-e55c36465218 28edb445-2abd-4cba-9143-e55c36465218 Tue, 22 Sep 2026 11:06:37 GMT budget OpenRouter model catalogue OpenRouter raised the completion token price for deepseek/deepseek-v4-flash-latest from $0.0000008 to $0.000001 per token, a 25% increase. Prompt pricing ($0.00000003) and context length (1.31M tokens) remain unchanged. Why it matters: If you generate high volumes of completion tokens with this model, your monthly API costs will increase by 25% at current usage levels. A workload generating 1 billion completion tokens per month now costs $1,000 instead of $800. What to do: Pull your last 30 days of completion token usage for this model ID from OpenRouter logs or your own metrics, multiply by the new rate, and update your cost forecast. If the increase is material, evaluate whether switching to the non-latest pinned version or a different model makes sense. https://radar.trulysovereign.ai/items/dde3cc56-c10e-4d7b-a1d5-540688714a66 dde3cc56-c10e-4d7b-a1d5-540688714a66 Tue, 22 Sep 2026 11:06:28 GMT budget OpenRouter model catalogue OpenRouter's deepseek-pro-latest model pricing changed: prompt tokens from $0.000000558624 to $0.000000624 per token (+11.7%), completion tokens from $0.000001675872 to $0.00000288 per token (+71.6%). Max completion tokens increased from 384,000 to 943,718. Why it matters: Completion token costs rose 72%, which compounds on long-form generation workloads. A 10,000-token completion that cost $0.0168 now costs $0.0288, a $0.012 increase per call. High-volume applications will see material budget impact. What to do: Pull last 30 days of deepseek-pro-latest usage from OpenRouter logs, calculate new cost at current volume, and update your monthly forecast. If completion length averages above 5,000 tokens, evaluate whether the increased max_completion_tokens justifies the 72% cost increase or whether to switch models. https://radar.trulysovereign.ai/items/4c917c6c-75bf-45df-8c0b-4e990a37c4ff 4c917c6c-75bf-45df-8c0b-4e990a37c4ff Tue, 22 Sep 2026 11:06:19 GMT verify_only OpenRouter model catalogue The top_provider.max_completion_tokens value increased from 32768 to 235929 tokens. Context length remains 262144 tokens. Pricing unchanged at $0.00000015 per prompt token and $0.0000012 per completion token. Why it matters: Applications currently capped at 32768 output tokens can now request up to 235929 tokens per completion, enabling longer-form generation without chunking. Existing requests under the old limit continue to work unchanged. What to do: If you have workflows that chunk outputs to stay under 32768 tokens, test whether single-pass generation at higher token counts improves quality or reduces latency for your use case. https://radar.trulysovereign.ai/items/e4280679-fc5a-40d0-9f2a-fcd8bd326d2e e4280679-fc5a-40d0-9f2a-fcd8bd326d2e Tue, 22 Sep 2026 03:48:19 GMT verify_only Claude Platform release notes The thinking-binding-controls-2026-08-01 beta header adds a thinking_mismatch_allowed entry type to input_transformations responses. This flags thinking blocks that fail prefix validation but are not rejected—specifically on Claude Fable 5.1 requests from accounts created before August 31, 2026, when prefix_mismatch_behavior is unset. The flagged blocks still reach the model unchanged. Why it matters: If you edit conversation history containing thinking blocks, the model may receive inconsistent context without your knowledge. This header lets you detect those edits in production traffic before enforcement becomes mandatory for your account, preventing silent context corruption that could degrade model behavior. What to do: Add the thinking-binding-controls-2026-08-01 header to requests on Claude Fable 5.1, log all thinking_mismatch_allowed entries for one week, and audit any history-editing code paths that trigger them. If you find mismatches, either stop editing thinking blocks or set prefix_mismatch_behavior explicitly before the August 31 cutoff applies to your account. https://radar.trulysovereign.ai/items/b97fd067-22c1-4490-a906-36a955902cdb b97fd067-22c1-4490-a906-36a955902cdb Mon, 21 Sep 2026 23:45:06 GMT budget Anthropic pricing The Individual pricing tier no longer exists; it has been renamed to Team. Claude Code, previously included in both Pro ($17/month) and Max (from $100/month), is now only included in the Max tier. The Pro tier description no longer lists Claude Code as a feature. Why it matters: If you are on the Pro tier and rely on Claude Code for development workflows, you will need to upgrade to Max (minimum $100/month, a $83/month increase) to retain access. The tier rename suggests Anthropic is repositioning the product away from individual use cases. What to do: Check your current subscription tier and whether your team uses Claude Code. If you are on Pro and depend on Claude Code, budget for the upgrade to Max or identify alternative tooling before your next billing cycle. https://radar.trulysovereign.ai/items/f1aa4b15-ecac-4178-aff6-668bee4c188d f1aa4b15-ecac-4178-aff6-668bee4c188d Mon, 21 Sep 2026 23:44:57 GMT budget Anthropic pricing The Individual pricing tier no longer exists; it has been renamed to Team. Claude Code, previously included in both Pro ($17/month) and Max (from $100/month), is now only included in the Max tier. The Pro tier description no longer lists Claude Code as a feature. Why it matters: If you are on the Pro tier and rely on Claude Code for development workflows, you will need to upgrade to Max (minimum $100/month, a $83/month increase) to retain access. The tier rename suggests Anthropic is repositioning the product away from individual use cases. What to do: Check your current subscription tier and whether your team uses Claude Code. If you are on Pro and depend on Claude Code, budget for the upgrade to Max or identify alternative tooling before your next billing cycle. https://radar.trulysovereign.ai/items/2c3b5ee0-f4fe-4aad-9140-fed1cbb187d1 2c3b5ee0-f4fe-4aad-9140-fed1cbb187d1 Mon, 21 Sep 2026 23:44:47 GMT budget Anthropic pricing The Individual pricing tier no longer exists; it has been renamed to Team. Claude Code, previously included in both Pro ($17/month) and Max (from $100/month), is now only included in the Max tier. The Pro tier description no longer lists Claude Code as a feature. Why it matters: If you are on the Pro tier and rely on Claude Code for development workflows, you will need to upgrade to Max (minimum $100/month, a $83/month increase) to retain access. The tier rename suggests Anthropic is repositioning away from individual users toward team-oriented pricing. What to do: Check your current Anthropic subscription tier and confirm whether you use Claude Code. If you are on Pro and depend on Claude Code, budget for the upgrade to Max or identify alternative tooling before your next billing cycle. https://radar.trulysovereign.ai/items/d680fb6d-7663-4909-9139-8ded8a4e7a4c d680fb6d-7663-4909-9139-8ded8a4e7a4c Mon, 21 Sep 2026 23:44:37 GMT budget Anthropic pricing The Individual pricing tier no longer exists; it has been renamed to Team. Claude Code, previously included in both Pro ($17/month) and Max (from $100/month), is now only included in the Max tier. The Pro tier description no longer lists Claude Code as a feature. Why it matters: If you are on the Pro tier and rely on Claude Code for development workflows, you will need to upgrade to Max (minimum $100/month, a $83/month increase) to retain access. The tier rename suggests Anthropic is repositioning its offering away from individual users toward team-based pricing. What to do: Check your current Anthropic subscription tier and confirm whether you use Claude Code. If you are on Pro and depend on Claude Code, budget for the upgrade to Max or identify alternative tooling before your next billing cycle. https://radar.trulysovereign.ai/items/26e9939b-7751-4811-b138-460f1ce62c76 26e9939b-7751-4811-b138-460f1ce62c76 Mon, 21 Sep 2026 23:44:10 GMT budget Anthropic pricing The Individual pricing tier no longer exists; it has been renamed to Team. Claude Code, previously included in both Pro ($17/month) and Max (from $100/month), is now only available in the Max tier. The Pro tier description now lists Claude Code as a feature, but the Max tier no longer states 'Claude Code included' in its header. Why it matters: If you are on the Pro tier and use Claude Code, you will need to upgrade to Max (minimum $100/month, a 5.9x increase) to retain access. If you are on what was called the Individual tier, your tier name has changed but feature access appears unchanged. What to do: Check your current Anthropic subscription tier and confirm whether you use Claude Code. If you are on Pro and depend on Claude Code, budget for the upgrade to Max or identify alternative tooling before the change takes effect. https://radar.trulysovereign.ai/items/68b7879f-1277-4f33-a8f6-6843ca33d56e 68b7879f-1277-4f33-a8f6-6843ca33d56e Mon, 21 Sep 2026 23:43:42 GMT verify_only Azure Foundry model retirement schedule gpt-realtime-2.1 and gpt-realtime-2.1-mini changed status from Preview to GA effective 2026-07-07. Both models have a retirement date of 2027-06-25. Why it matters: GA status means these models now carry production SLA guarantees and support commitments that Preview versions do not. Any production workloads currently using these models are now covered under standard Azure support terms. What to do: If you are using either model in production, verify your deployment is receiving GA-level support and SLA coverage. If you were blocking production use pending GA status, that gate is now cleared. https://radar.trulysovereign.ai/items/34391a29-307f-44c4-b6ca-5c9775db7859 34391a29-307f-44c4-b6ca-5c9775db7859 Mon, 21 Sep 2026 23:43:33 GMT verify_only Anthropic pricing The 'Individual' and 'Developer' tier labels were removed from the pricing page. Claude Code, previously listed only under Pro, now appears under both Pro ($17/month) and Max (from $100/month). The Max tier description changed from 'For people who work with Claude all day' to the same phrasing. Why it matters: If you are on a legacy 'Individual' or 'Developer' plan, the tier name change may affect how you reference your subscription in billing systems or support requests. The Claude Code availability shift does not change what Pro users have access to, but clarifies that Max users also get it—relevant if you were evaluating Max specifically for coding features. What to do: Check your Anthropic account dashboard to confirm your current tier name matches the new structure. If you have billing automation or internal documentation referencing 'Individual' or 'Developer' tiers, update those labels to match the current naming. https://radar.trulysovereign.ai/items/da960a09-8d62-4cd9-aa80-fc01b818f5fc da960a09-8d62-4cd9-aa80-fc01b818f5fc Mon, 21 Sep 2026 23:43:23 GMT migrate Azure Foundry model retirement schedule The retirement date for gpt-4.1-nano moved from 2026-10-14 to 2027-04-14, a six-month extension. The model entered Legacy status on 2025-04-14. Why it matters: If you are running gpt-4.1-nano in production, you now have until mid-April 2027 to complete migration instead of mid-October 2026. This changes the timeline for any scheduled replacement work. What to do: Check your deployment manifests and infrastructure-as-code for references to gpt-4.1-nano. If found, update your migration plan to reflect the new 2027-04-14 deadline and communicate the revised timeline to stakeholders. https://radar.trulysovereign.ai/items/9f8a0f68-d04b-4beb-ba71-cce82d816d9b 9f8a0f68-d04b-4beb-ba71-cce82d816d9b Mon, 21 Sep 2026 23:43:07 GMT budget Anthropic pricing Max 5x and Max 20x tiers now offer monthly billing only, down from 'Monthly and annual'. Free tier projects capped at 5, down from unlimited. Context window changed from fixed '200k' across all tiers to 'Up to 1M varies by model'. Claude Cowork feature removed entirely from all tiers. Why it matters: Teams on annual Max contracts cannot renew at the annual rate when their term expires, forcing monthly billing at potentially higher effective cost. Free tier users with more than 5 projects will lose access to older projects or need to upgrade. What to do: If you are on Max 5x or Max 20x with annual billing, check your renewal date and budget for monthly billing going forward. If you use Free tier with more than 5 projects, audit which projects contain work you need to preserve before the cap takes effect. https://radar.trulysovereign.ai/items/1025bd98-ec2e-4bba-a135-160375cd0aa8 1025bd98-ec2e-4bba-a135-160375cd0aa8 Mon, 21 Sep 2026 23:42:57 GMT migrate Azure Foundry model retirement schedule gpt-4.1-nano retirement moved from 2027-04-14 to 2026-10-14. gpt-image-1.5 retirement moved from 2027-06-16 to 2026-12-16. Both dates accelerated by six months. Why it matters: Any deployments using these models will stop working on the new dates. Six months less migration runway means work scheduled for Q2 2027 must now complete by Q4 2026. What to do: Audit Azure deployments for gpt-4.1-nano and gpt-image-1.5. If found, schedule migration to successor models before 2026-10-14 and 2026-12-16 respectively. https://radar.trulysovereign.ai/items/3cb22dae-6b0b-4297-8a3a-0124331f2357 3cb22dae-6b0b-4297-8a3a-0124331f2357 Mon, 21 Sep 2026 23:42:49 GMT verify_only Claude Platform release notes The Messages API now supports a `compaction` parameter (beta header `compact-2026-09-04`) that returns a signed compaction block summarizing sent messages. This block replaces the original messages in subsequent requests while preserving recent turns verbatim. Why it matters: Long-running conversations currently hit context limits or incur cumulative token costs. This allows you to compress conversation history server-side while keeping recent exchanges intact, reducing both token usage and request size without losing conversation continuity. What to do: Test the beta header `compact-2026-09-04` in a non-production environment with one long-running conversation flow. Measure token reduction and verify that responses remain coherent when using the compaction block instead of full message history. https://radar.trulysovereign.ai/items/e0890059-8377-432d-a8bb-375aa1351c22 e0890059-8377-432d-a8bb-375aa1351c22 Sat, 12 Sep 2026 10:13:50 GMT migrate OpenAI deprecations OpenAI announced on September 11, 2026 that gpt-5.4-cyber is deprecated and will be removed from the API on October 1, 2026. The replacement model is gpt-5.6-cyber. Why it matters: Any code calling gpt-5.4-cyber will fail after October 1, 2026. You have 19 days to complete the migration. What to do: Grep your codebase for 'gpt-5.4-cyber', update all references to 'gpt-5.6-cyber', test the replacement model's output against your use cases, and deploy before October 1. https://radar.trulysovereign.ai/items/23bc0de3-09e4-4b40-8fef-f1254dbbec6d 23bc0de3-09e4-4b40-8fef-f1254dbbec6d Sat, 12 Sep 2026 03:31:37 GMT migrate Azure Foundry model retirement schedule o1-pro, o3, o3-deep-research, and o3-pro moved from GA to Deprecated status. All o1/o3 models now retire 2026-11-19 (previously scattered between 2026-10-01 and 2026-12-26). o3-mini and o4-mini successor changed from o4-mini to gpt-5.6-terra. Fine-tuning retirement for o4-mini shifted from 2027-04-16 to 2027-05-19. Why it matters: If you are running o1-pro, o3, o3-deep-research, or o3-pro in production, they are now deprecated with 9 weeks until retirement. The consolidated sunset date means you cannot stagger migrations. The successor model change means existing migration plans targeting o4-mini must be rewritten to target gpt-5.6-terra or gpt-5.6-sol. What to do: Grep your codebase for model identifiers o1-pro, o3, o3-deep-research, o3-pro, o3-mini, o4-mini. If any are in use, schedule migration to gpt-5.6-sol (for o1/o3 reasoning models) or gpt-5.6-terra (for o3-mini/o4-mini) before 2026-11-19. If you have fine-tuned o4-mini models, note the new training retirement date of 2027-05-19. https://radar.trulysovereign.ai/items/36c5e994-d0ff-4e5f-b145-fc5bb4119aa9 36c5e994-d0ff-4e5f-b145-fc5bb4119aa9 Fri, 11 Sep 2026 03:26:05 GMT verify_only Claude Platform release notes Claude Managed Agents now supports an 'auto' permission policy mode where the server evaluates each tool call and either runs it, denies it, or pauses for approval. The agent.tool_use and agent.mcp_tool_use events now include an 'evaluation' field reporting how each call was evaluated. The ant CLI adds 'ant beta:sessions connect' command to attach a terminal to live agent sessions, with an optional --web flag to serve the session viewer locally. Why it matters: If you are running Claude Managed Agents with custom permission policies, the new auto mode changes how tool call authorization works—the server now makes evaluation decisions instead of requiring you to implement all logic client-side. The new evaluation field in events means your logging or monitoring code may need updates to capture this data. The CLI command provides a new debugging path that bypasses your application's session handling. What to do: Review your agent permission policy configurations and decide whether to adopt auto mode or continue with existing policies. Update any event handlers that parse agent.tool_use or agent.mcp_tool_use events to handle the new evaluation field. Test the ant beta:sessions connect command in development to understand how it interacts with your session management. https://radar.trulysovereign.ai/items/a6e8fa9f-1df0-4980-b269-29cd94f20820 a6e8fa9f-1df0-4980-b269-29cd94f20820 Thu, 10 Sep 2026 20:30:21 GMT verify_only Azure Foundry model retirement schedule gpt-image-2.5-flare and gpt-image-2.5-sunburst reached GA on 9 September 2026 and are scheduled for retirement on 9 September 2027. Both models now appear on the Azure Foundry model retirement schedule. Why it matters: If you adopt either model after GA, you have exactly one year before forced migration. The retirement date is fixed at announcement, which is unusually short for a GA model lifecycle. What to do: Check your image generation endpoints for references to gpt-image-2.5-flare or gpt-image-2.5-sunburst. If you are evaluating these models, factor the 9 September 2027 cutoff into your adoption decision now. https://radar.trulysovereign.ai/items/43a0143c-01bf-406b-925f-3e8c8b0dc3e4 43a0143c-01bf-406b-925f-3e8c8b0dc3e4 Wed, 09 Sep 2026 20:34:35 GMT verify_only Claude Platform release notes Per-message effort changes (beta) are now supported on Google Cloud for Claude Fable 5.1, Claude Mythos 5.1, and Claude Opus 5. This requires the mid-conversation-output-config-2026-07-01 beta header, matching the existing implementation on other platforms. Why it matters: If you run Claude on Google Cloud and want to adjust model effort mid-conversation (for example, switching between speed and quality), you can now do so without changing platforms. Previously this beta feature was unavailable on Google Cloud deployments. What to do: If you use Claude on Google Cloud and want per-message effort control, test the mid-conversation-output-config-2026-07-01 header with one of the three supported models in a non-production environment to verify the behavior matches your use case. https://radar.trulysovereign.ai/items/69352f03-8716-45ab-9e08-cbd52906df4d 69352f03-8716-45ab-9e08-cbd52906df4d Fri, 04 Sep 2026 20:23:26 GMT verify_only Anthropic pricing The pricing comparison table no longer lists 'Access to Research' as a Pro tier feature. The row showing 'Research' with 'No' for Free tier has been removed entirely. Why it matters: If your team or users rely on Research access through Pro subscriptions, this removal suggests the feature may no longer be available at that tier, potentially requiring a plan upgrade or eliminating access entirely. What to do: Log into an active Pro account and verify whether Research is still accessible in the product. If it is missing, contact Anthropic support to confirm whether this is a pricing page error or an actual feature removal, and determine what tier now includes Research. https://radar.trulysovereign.ai/items/50947ff5-1d48-4f0e-b695-daf7e44eec1e 50947ff5-1d48-4f0e-b695-daf7e44eec1e Thu, 03 Sep 2026 20:37:01 GMT verify_only Anthropic pricing The pricing page no longer displays Team plan pricing ($20-$25/seat standard, $100-$125/seat premium) or Enterprise plan details ($20/seat + API usage). The Education plan section was removed entirely. API pricing tables for all models (Fable 5.1, Opus 5, Sonnet 5, Haiku 4.5) and legacy models are no longer visible in the public diff. The feature comparison now shows only Free, Pro, Max 5x, and Max 20x tiers. Why it matters: If you are evaluating Team or Enterprise plans, you can no longer see pricing or feature details without contacting sales. If you are budgeting for multi-seat deployments or comparing API costs across models, the reference data is no longer publicly available. Existing customers are unaffected, but new procurement requires a sales conversation where self-serve comparison was previously possible. What to do: If you have an active RFP or budget proposal referencing the removed pricing, screenshot your current contract or contact Anthropic sales to confirm whether published rates still apply. If you are comparing API costs, check whether you have the model pricing cached locally or in your cost model; if not, request a rate card from your account team before the next planning cycle. https://radar.trulysovereign.ai/items/77b25006-adfb-4e7d-8b6f-1fe2cf626027 77b25006-adfb-4e7d-8b6f-1fe2cf626027 Thu, 03 Sep 2026 10:48:13 GMT verify_only Claude Platform release notes As of September 1, 2026, audio files generated by Claude Fable 5.1 and Claude Mythos 5.1 through the code execution tool now carry C2PA Content Credentials when retrieved via the Files API. Previously only image and video files were marked. Why it matters: If your application processes or stores audio artifacts from Claude's code execution, those files now contain embedded provenance metadata. Systems that validate file integrity, strip metadata, or assume clean audio output may reject or mishandle these files. What to do: Test your audio file handling pipeline with code-execution-generated audio from Fable 5.1 or Mythos 5.1. Verify that parsers, transcoders, and storage systems accept C2PA-marked audio without errors or unexpected metadata stripping. https://radar.trulysovereign.ai/items/c5855813-268c-4213-8f94-7cf3f38ef45c c5855813-268c-4213-8f94-7cf3f38ef45c Thu, 03 Sep 2026 03:20:06 GMT migrate Azure Foundry model retirement schedule Azure Foundry added a retirement date of 2027-12-05 for the claude-fable-5-1 model. The model is currently in Preview status. Why it matters: Any code or configuration using claude-fable-5-1 by name will stop working after that date. Preview models typically have shorter lifespans than GA releases, and 15 months is the notice window. What to do: Grep your codebase and infrastructure-as-code for 'claude-fable-5-1' references. If found, add a calendar reminder for Q3 2027 to migrate to whatever replaces it, or pin to claude-fable-5 if that model remains supported. https://radar.trulysovereign.ai/items/0ee7793d-d108-4471-9679-5872fdd2c0ee 0ee7793d-d108-4471-9679-5872fdd2c0ee Thu, 03 Sep 2026 03:19:57 GMT migrate Claude Platform release notes The Admin API (user management and spend limits), Analytics API, and Compliance API now require the anthropic-version header on every request, matching the existing requirement across the rest of the Claude API. Documentation updated September 1, 2026. Why it matters: Requests to these three Enterprise endpoints will fail without the header. If your integration calls these endpoints and does not currently send anthropic-version, those calls are now broken. What to do: Grep your codebase for calls to Admin, Analytics, and Compliance API endpoints. Add the anthropic-version header to each request if missing. Test against production to confirm the calls succeed. https://radar.trulysovereign.ai/items/d3535870-6137-4cc1-92e4-3852981b3a3d d3535870-6137-4cc1-92e4-3852981b3a3d Tue, 01 Sep 2026 20:40:45 GMT verify_only Anthropic pricing Anthropic released Fable 5.1, replacing Fable 5 in the current model lineup. Prompt cache read pricing dropped from $1/MTok to $0.25/MTok. Fable 5 moved to legacy pricing with its original $1/MTok cache read rate. Input ($10/MTok) and output ($50/MTok) pricing unchanged. Why it matters: If you use Fable with prompt caching for long-context or repeated-prompt workloads, cache read costs drop 75% on the new version. Staying on Fable 5 means paying 4x more for cache reads than the current model. The version bump may also include capability changes not reflected in pricing. What to do: Check your API calls: if you specify 'fable-5' explicitly, test whether that identifier now resolves to 5.1 or stays pinned to legacy 5. If unpinned, you get the cost reduction automatically. If pinned, evaluate 5.1 for your use case and update the model identifier to capture the savings. https://radar.trulysovereign.ai/items/cf2722ee-e40f-454a-bc92-5775238664a9 cf2722ee-e40f-454a-bc92-5775238664a9 Tue, 01 Sep 2026 20:40:33 GMT migrate Azure Foundry model retirement schedule claude-opus-4-1 marked Retired with effective date 2026-08-05. claude-opus-5 added to supported models list. Two new preview models added: claude-fable-5-1 and claude-mythos-5-1. Why it matters: Any code or configuration explicitly requesting claude-opus-4-1 will fail after August 5, 2026. If you are using version pinning or explicit model identifiers rather than latest-stable aliases, calls will return errors once the retirement date passes. What to do: Grep your codebase and infrastructure-as-code for 'claude-opus-4-1' and replace with 'claude-opus-5'. Test the replacement in staging before August 5, 2026. https://radar.trulysovereign.ai/items/ba917d39-270a-4508-8b5e-ddc025e3fd7c ba917d39-270a-4508-8b5e-ddc025e3fd7c Tue, 01 Sep 2026 20:40:24 GMT migrate Claude Platform release notes Claude Fable 5.1 and Mythos 5.1 launched September 1, 2026. tool_choice types 'any' and 'tool' now return 400 errors; only 'auto' and 'none' work. Cache read pricing dropped from 0.1x to 0.025x base price ($0.25/MTok). Both models enforce 30-day minimum data retention with no zero-retention option unless Anthropic authorizes it. New accounts created August 31, 2026 or later get 400 errors when replaying thinking blocks after system prompt, tools, or message history changes. Why it matters: If you use tool_choice='any' or tool_choice='tool' in production, requests to Fable 5.1 will fail immediately. If your contract or compliance posture requires zero data retention, you cannot use these models without explicit authorization. The thinking block validation will break conversation replay for new accounts if you modify context mid-conversation. What to do: Grep your codebase for tool_choice parameters set to 'any' or 'tool' and migrate to structured outputs or strict tool use before switching to Fable 5.1. Verify your data retention agreement permits 30-day retention or request authorization from Anthropic if you need zero retention. If you create new accounts after August 31, audit any code that replays thinking blocks after modifying system prompts or tools. https://radar.trulysovereign.ai/items/2caf39c0-b116-452d-8c4f-e5c6119c3081 2caf39c0-b116-452d-8c4f-e5c6119c3081 Mon, 31 Aug 2026 04:10:47 GMT migrate OpenAI deprecations On August 26, 2026, OpenAI announced that whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize will be removed from the API on February 26, 2027. Recommended replacements are gpt-live-transcribe or gpt-transcribe. Why it matters: Any application using these four models for audio transcription will break on February 26, 2027 unless migrated to the replacement models. You have six months to test and deploy the migration. What to do: Grep your codebase for 'whisper-1', 'gpt-4o-transcribe', 'gpt-4o-mini-transcribe', and 'gpt-4o-transcribe-diarize'. For each usage, schedule migration to gpt-live-transcribe or gpt-transcribe before February 26, 2027. https://radar.trulysovereign.ai/items/c00829cd-4938-43d2-9f3f-f7eaa759fb5b c00829cd-4938-43d2-9f3f-f7eaa759fb5b Sat, 29 Aug 2026 05:59:09 GMT migrate Azure Foundry model retirement schedule MAI-Image-2.5, MAI-Image-2.5-Flash, and MAI-Image-2.5-Pro added to Azure Foundry retirement schedule. All three entered Preview on June 2 or June 19, 2026, and will be retired October 1, 2026. MAI-Image-2e status changed from Preview to Retired. Why it matters: If you are using or planning to use any MAI-Image-2.5 variant, you have a four-month window from preview to retirement. Code written against these models will break on October 1, 2026 unless migrated to a successor model. What to do: Grep your codebase for 'MAI-Image-2.5', 'MAI-Image-2.5-Flash', and 'MAI-Image-2.5-Pro'. If found, schedule migration work to complete before October 1, 2026. If not using these models, no action required. https://radar.trulysovereign.ai/items/f0d559ef-a8a6-4d04-adbd-3a1e98096d5d f0d559ef-a8a6-4d04-adbd-3a1e98096d5d Fri, 28 Aug 2026 21:46:44 GMT verify_only Azure Foundry model retirement schedule Azure Foundry now offers Claude models in two versions: Version 1 (Anthropic infrastructure) and Version 2 (Azure infrastructure). Multiple models moved from Preview to GA status. claude-opus-4-1 was removed from the schedule (previously marked Retired 2026-08-05). New model claude-fable-5 added with retirement date 2027-12-05. claude-haiku-4-5 Preview variant removed; GA variants now exist for both Version 1 and Version 2, both retiring 2026-10-19. Why it matters: If you are calling Claude models by name without specifying version, your requests may now route to different infrastructure (Anthropic vs Azure) with potentially different latency, availability, or billing characteristics. The removal of claude-opus-4-1 from the schedule means any code still referencing it is calling a model that no longer appears in the official retirement timeline. What to do: Audit your model identifiers to confirm which version (1 or 2) you are using. If you are calling claude-opus-4-1, verify whether it still responds or migrate to claude-opus-5 (GA, retires 2027-07-08). If you rely on claude-haiku-4-5, confirm whether you are on Version 1 or 2 and note the 2026-10-19 retirement date for both. https://radar.trulysovereign.ai/items/2769b47b-ace9-4cdc-a8ba-fac335be0ded 2769b47b-ace9-4cdc-a8ba-fac335be0ded Fri, 28 Aug 2026 21:46:30 GMT verify_only Claude Platform release notes Compliance API session endpoints for Cowork and Claude Code are now GA (out of beta). Local session endpoints now return transcripts for Claude Science sessions (product_surface: claude_science) and Claude for Microsoft 365 sessions in Excel, PowerPoint, Word, and Outlook (product_surface values starting with office_agents), both in beta for Enterprise organizations. No new authentication required—existing Compliance Access Key with read:compliance_user_data scope works. Why it matters: If you retrieve compliance transcripts, your existing integration now captures two additional product surfaces without code changes. If you filter or route transcripts by product_surface value, unrecognized values will appear in your logs or break strict validation. What to do: Check your Compliance API response handling for product_surface filtering or validation. Add claude_science and office_agents* to your expected values list if you whitelist them, or confirm your code handles unknown values gracefully. https://radar.trulysovereign.ai/items/11e6f928-cc7f-4d8f-8d61-0dd49e322651 11e6f928-cc7f-4d8f-8d61-0dd49e322651 Wed, 26 Aug 2026 19:49:45 GMT verify_only Claude Platform release notes The computer_toolset_20260801 and browser_toolset_20260801 toolsets are now available on Google Cloud for Claude Fable 5, Claude Mythos 5, Claude Opus 5, Claude Sonnet 5, and Claude Opus 4.8. The tools entries format matches the Claude API. Why it matters: If you run Claude on Google Cloud and have been waiting for computer use or browser use capabilities, you can now deploy those features without migrating to a different cloud provider or API endpoint. What to do: If you have Google Cloud deployments that need computer use or browser use, test the toolsets with your existing tools entries format—no API changes are required from the Claude API implementation. https://radar.trulysovereign.ai/items/45db7e68-2aed-4776-bb75-8dc35012c6f1 45db7e68-2aed-4776-bb75-8dc35012c6f1 Tue, 25 Aug 2026 06:49:24 GMT verify_only Azure Foundry model retirement schedule The retirement date for the sora-2 preview model moved from 15 September 2026 to 15 October 2026. Why it matters: If you planned migration work around the September deadline, you now have four additional weeks. If you already migrated, the extension changes nothing. What to do: Check your migration schedule for sora-2. If you have not started and the September date was tight, you can defer by up to four weeks. If you already migrated or the date was not a constraint, no action needed. https://radar.trulysovereign.ai/items/e503476f-0504-4f5a-861a-aec1994696ea e503476f-0504-4f5a-861a-aec1994696ea Sat, 22 Aug 2026 01:15:54 GMT pin_or_freeze Claude Platform release notes The Messages API now defaults top_p to 0.99 instead of 0.999 across all models, effective immediately. Extended thinking calls can now set top_p between 0.95 and 1.0, expanded from the previous narrower range. Why it matters: Every request not explicitly setting top_p now uses different sampling behavior, which changes response variability and may affect tool call reliability, output consistency in production workflows, and any testing or validation tied to specific model behavior. What to do: Audit your Claude API calls: if you rely on the previous default behavior, explicitly set top_p: 0.999 in all requests. If you use extended thinking and want the old constraint, pin top_p below 0.95 or above 1.0 will now error—test your parameter validation. https://radar.trulysovereign.ai/items/cb16fe1f-0d67-4e11-88c8-1290ac610dac cb16fe1f-0d67-4e11-88c8-1290ac610dac Sat, 22 Aug 2026 01:15:30 GMT migrate Claude Platform release notes Released August 20, 2026. Python SDK v1.0 removes the legacy Text Completions API, the temperature/top_p/top_k parameters on Messages methods, and tool runner's client-side compaction_control. Minimum Python version is now 3.10. HTTP layer switches from httpx to httpx2. Async client .with_raw_response results now require await response.parse(). AnthropicBedrock no longer defaults to us-east-1 when AWS region is unconfigured. Why it matters: Code using the removed Text Completions API or deprecated parameters will break immediately on upgrade. Applications running Python 3.9 or earlier cannot install this version. Async response handling that omits await response.parse() will fail. AnthropicBedrock calls without explicit region configuration will error instead of silently using us-east-1. What to do: Before upgrading: grep your codebase for 'text_completions', 'temperature', 'top_p', 'top_k' in Messages calls, and 'compaction_control'; verify Python version is 3.10+; check all .with_raw_response usage includes await response.parse(); confirm AnthropicBedrock has explicit region set. Consult the v1 migration guide for replacement patterns. https://radar.trulysovereign.ai/items/ce14b9ea-b4c1-48bd-9827-60602c108688 ce14b9ea-b4c1-48bd-9827-60602c108688 Sat, 22 Aug 2026 01:15:19 GMT verify_only Claude Platform release notes The max-tokens-3-5-sonnet-2024-07-15 header is no longer required to access 8,192-token outputs on Claude Sonnet 3.5. The feature moved from beta to general availability on August 19, 2024. Why it matters: Code that sets this header will continue to work but is now redundant. Teams that avoided the feature due to beta status can now use extended outputs without special headers. What to do: Search your codebase for 'max-tokens-3-5-sonnet-2024-07-15' and remove the header from API calls. The functionality remains identical without it. https://radar.trulysovereign.ai/items/ae29ba19-50c2-440b-8b83-04ca0bf1aaf0 ae29ba19-50c2-440b-8b83-04ca0bf1aaf0 Fri, 21 Aug 2026 01:19:37 GMT migrate Claude Platform release notes The computer use tool is now GA as computer_toolset_20260801, replacing earlier beta versions. The upgrade changes the request shape and tool handling. The Files API is GA with a new response format: file expiration (expires_in_seconds on upload, expires_at in responses) and page/next_page pagination with ids[] filter. Both changes take effect immediately when you stop sending the beta headers. Why it matters: Existing integrations using the beta computer use tool will break if you switch to the GA version without updating your request structure. Files API clients that drop the beta header will receive a different response format and pagination scheme, breaking any code that expects the old structure. What to do: If you use the computer use tool, read the migration guide at 'Migrate from computer_20251124' before removing the beta header. If you use the Files API, test your pagination and expiration handling against the GA format before dropping the files-api-2025-04-14 header from production requests. https://radar.trulysovereign.ai/items/205c4cd5-1cf4-4e6c-8386-ef3c97c4a0a5 205c4cd5-1cf4-4e6c-8386-ef3c97c4a0a5 Thu, 20 Aug 2026 18:39:57 GMT migrate Claude Platform release notes Effective August 19, 2026, three Claude APIs moved from beta to general availability: Files API (/v1/files), Skills API (/v1/skills), and Admin API user-management endpoints. Beta headers (files-api-2025-04-14, skills-2025-10-02, anthropic-beta: ce-user-management-2026-07-13) are no longer required. The Files API GA response format differs from beta: pagination changed from page/next_page to ids[] filter, and file expiration now uses expires_in_seconds (upload) and expires_at (response) instead of the beta format. Why it matters: Code sending beta headers continues to work but receives the old response format. Code that stopped sending headers or never sent them now receives the GA format with different pagination and expiration field names, which will break parsing logic expecting the beta structure. What to do: Grep your codebase for 'files-api-2025-04-14', 'skills-2025-10-02', and 'ce-user-management-2026-07-13'. For Files API calls, verify whether your code parses expires_at and handles ids[] pagination, or if it expects the beta format. If you need the beta format, confirm the header is still being sent; if you need GA format, remove the header and update response parsing. https://radar.trulysovereign.ai/items/ebad5c37-05e6-4d61-96a7-661d9006c8d5 ebad5c37-05e6-4d61-96a7-661d9006c8d5 Thu, 20 Aug 2026 01:15:58 GMT migrate Azure Foundry model retirement schedule Azure Foundry moved gpt-4.1-nano retirement from 2026-10-14 to 2027-04-14. The model's status label changed from 'Deprecated' to 'Legacy' on 2025-04-14. Three other models (gpt-4.1, gpt-4.1-mini, gpt-4o) also changed to 'Legacy' status with 2027-04-14 retirement dates. Why it matters: If you planned migration work around the October 2026 deadline, you now have six additional months. Teams that deprioritized gpt-4.1-nano migration because of the tight timeline can reschedule without urgency. What to do: Check your migration backlog for gpt-4.1-nano work scheduled before 2027-04-14. If you deferred it due to the October deadline, update the target date and reprioritize against other work. https://radar.trulysovereign.ai/items/2a3739cf-3ab9-49a7-ab94-0249afd03331 2a3739cf-3ab9-49a7-ab94-0249afd03331 Wed, 19 Aug 2026 18:36:14 GMT migrate Azure Foundry model retirement schedule Azure added DeepSeek-V4-Flash-0731 to the retirement schedule. The model enters GA on 2026-07-31 and preview access ends on 2026-12-03. Why it matters: If you are testing this model in preview, you have until 2026-12-03 to migrate to the GA version or switch models. Preview endpoints will stop working after that date. What to do: Check if any code or configs reference DeepSeek-V4-Flash-0731 in preview. If so, schedule migration to the GA version before 2026-12-03. https://radar.trulysovereign.ai/items/1a010efc-3b68-4ca7-909c-3aa558516e6b 1a010efc-3b68-4ca7-909c-3aa558516e6b Tue, 18 Aug 2026 01:15:03 GMT migrate Azure Foundry model retirement schedule DeepSeek-R1, MAI-Image-2, Stable-Diffusion-3.5-Large, Stable-Image-Core, and Stable-Image-Ultra moved from Legacy/Preview/Deprecated to Retired status between 31 July and 15 August 2026. MAI-Transcribe-1 added to retirement schedule with cutoff date 15 September 2026. MAI-Transcribe-1.5 added as successor model. Why it matters: Retired models are no longer callable via API. Any code still referencing DeepSeek-R1, MAI-Image-2, or the three Stable models will fail. MAI-Transcribe-1 stops working in 28 days. What to do: Grep your codebase for 'DeepSeek-R1', 'MAI-Image-2', 'Stable-Diffusion-3.5-Large', 'Stable-Image-Core', 'Stable-Image-Ultra', and 'MAI-Transcribe-1'. Remove or replace references before 15 September 2026 for MAI-Transcribe-1; the other five are already dead. https://radar.trulysovereign.ai/items/dee30267-556c-4501-82bf-63d8d2911111 dee30267-556c-4501-82bf-63d8d2911111 Mon, 17 Aug 2026 00:11:56 GMT verify_only Claude Platform release notes The Claude API now returns an anthropic-workspace-id header in all responses, containing the wrkspc_-prefixed workspace ID that the request's API key or access token resolved to, including the organization's Default Workspace. This applies to all API requests immediately. Why it matters: If you parse API responses or log headers for debugging, this new header will appear in all responses. Multi-workspace organizations can now programmatically identify which workspace handled each request without inferring from the API key alone. What to do: Check whether your HTTP client or logging pipeline treats unexpected headers as errors. If you operate multiple workspaces, decide whether to log this header for request attribution. https://radar.trulysovereign.ai/items/9cf1aa5e-9153-4e85-9f67-36dd983b53f9 9cf1aa5e-9153-4e85-9f67-36dd983b53f9 Mon, 17 Aug 2026 00:11:48 GMT verify_only Claude Platform release notes Three new Compliance API endpoints launched August 11, 2026 in beta for Enterprise organizations: GET /v1/compliance/apps/sessions/local lists all local sessions, GET /v1/compliance/apps/sessions/local/{session_id} retrieves session metadata, and GET /v1/compliance/apps/sessions/local/{session_id}/messages returns the full transcript. These cover Cowork and Claude Code sessions running on user machines. Authentication uses existing Compliance Access Keys with read:compliance_user_data scope. Why it matters: Your compliance tooling can now audit conversations that happen in desktop applications, not just API calls. If you already pull compliance data for governance or retention, these sessions were previously invisible and are now retrievable through the same authentication mechanism. What to do: If you run compliance exports or audits, add these three endpoints to your retrieval jobs. Test with GET /v1/compliance/apps/sessions/local using your existing Compliance Access Key to confirm scope and verify the transcript format matches your storage schema. https://radar.trulysovereign.ai/items/265d9891-c114-4e7c-a001-eac1b136629f 265d9891-c114-4e7c-a001-eac1b136629f Mon, 17 Aug 2026 00:11:37 GMT budget Claude Platform release notes Claude Sonnet 5 pricing remains at $2 input / $10 output per MTok indefinitely. The previously scheduled increase to $3 / $15 per MTok on September 1, 2026 will not occur, announced August 10, 2026. Why it matters: If you budgeted for the September price increase or delayed Sonnet 5 adoption expecting higher costs, your forecast is now wrong by 33-50% depending on input/output mix. Teams that pinned to older models to avoid the increase can now migrate without the cost penalty. What to do: Reforecast Claude Sonnet 5 spend using $2/$10 rates with no September step-up. If you delayed migration from Sonnet 4.6 due to the planned increase, re-evaluate the timeline—the cost barrier is gone.