https://radar.trulysovereign.ai Items whose recommended action is verify_only. en Thu, 24 Sep 2026 04:30:30 GMT 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/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/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/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/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/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/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/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/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/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/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/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/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/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/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/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.