Daily digest · 2026-10-25 (UTC)

What broke for AI agents on 25 Oct 2026

2 incidents on 1 status page.

Final. Built 2026-10-31 11:03 UTC, 7 hours after the day ended; this page will not change.

What broke

Anthropic: “Elevated errors for Claude Fable 5.1”major impact · opened 08:17 UTC · resolved 08:49 UTC (32 minutes) · Claude API (api.anthropic.com) | Claude Code | Claude Cowork | claude.aistatus page ↗provider-incidents v62
Anthropic: “Elevated errors for Claude Sonnet 5.5 and other models”major impact · opened 06:14 UTC · resolved 07:53 UTC (99 minutes) · Claude API (api.anthropic.com) | Claude Code | Claude Cowork | claude.aistatus page ↗provider-incidents v61

No new major version of the SDKs and tools we follow.

8 other releases on the day

@mistralai/mistralai 3.3.0 (npm)released 20:12 UTC · previous 3.2.0agent-sdk-releases v81
BerriAI/litellm v1.103.9published 08:57 UTC · previous v1.106.3release notes ↗agent-tool-releases v80
BerriAI/litellm v1.104.7published 09:12 UTC · previous v1.103.9release notes ↗agent-tool-releases v80
BerriAI/litellm v1.105.5published 09:12 UTC · previous v1.104.7release notes ↗agent-tool-releases v80
BerriAI/litellm v1.106.4published 09:39 UTC · previous v1.105.5release notes ↗agent-tool-releases v80
litellm 1.106.4 (pypi)released 09:32 UTC · previous 1.106.3agent-sdk-releases v80
mistralai 3.5.0 (pypi)released 19:53 UTC · previous 3.4.0agent-sdk-releases v81
modelcontextprotocol/servers v2.1.0published 00:03 UTC · previous v2.0.0release notes ↗agent-tool-releases v78

184 round-trips to 23 endpoints agents depend on, 0 failedmodel APIs, registries, chains, the x402 facilitator, status pages and two baselines · measured from macbook-kragent-api-observatory v104
500 of 500 remote MCP endpoints probed answered a handshake (100.0%), on 1 hosta different slice of the official registry each day, so the share is not compared across days · one attempt per endpoint, an auth challenge counts as an answer · measured from macbook-krmcp-server-liveness v24

Every line comes from one of WITAN's public datasets and links the version its records landed in: provider-incidents, agent-api-observatory, mcp-server-liveness, agent-sdk-releases, agent-tool-releases. They are free to read.
Incidents are what the providers' own status pages published (Anthropic, OpenAI, GitHub, Cloudflare, npm, Vercel), read every 6 hours; scheduled maintenance is left out. Follow a line's status page link for the provider's text.
Round-trips and MCP handshakes are WITAN's own measurements from one vantage point, named on each line. A failure there can be the network between us and the service; it is not a verdict on the service. A latency line needs the day's median at least 2× the previous 7 days' and 250 ms more.
Releases are those of the SDKs and agent tools the datasets follow; pre-releases are left out.
Days are UTC. A day is rebuilt every hour while it runs and frozen 7 hours after it ends. Model prices are not here: that dataset's collector is off. WITAN is a testnet preview run by one developer.