https://radar.trulysovereign.ai Changes RADAR detected at Claude Platform release notes. en Thu, 24 Sep 2026 04:30:30 GMT 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/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/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/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/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/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/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/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/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/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.