The one thing
01Claude Sonnet 5.5 is out at Sonnet 5 prices. Code written for Sonnet 5 can break on it in five documented ways, so change the model string only after checking them.
Anthropic launched claude-sonnet-5-5 on 28 September on its API, Bedrock, Google Cloud and Microsoft Foundry.1 List price is unchanged from Sonnet 5: $2 input, $10 output and $0.20 cache reads per million tokens. Anthropic says it costs up to 30% less per task because it uses fewer tokens.2 That per-task figure is Anthropic's own measurement.
The release notes list five breaks.1 thinking: {"type": "disabled"} must become {"type": "between_tools"}. Forced tool use (tool_choice any or tool) returns a 400. On the Claude API and Google Cloud the older computer_20251124 tool is refused. The advisor tool rejects Opus 4.8, Opus 4.7 and Sonnet 5 as advisors. Thinking blocks are tied to the model, the conversation and the account that produced them.
"disabled" thinking and forced tool_choice, and test any flow that moves a conversation between accounts.The account-binding change, and what GitHub measured
Thinking blocks from Sonnet 5.5 work only in the account that produced them or a linked one. When another account sends one, the API drops the block before the model sees it and the request still succeeds.1 Nothing errors, so a pipeline that hands conversations between accounts loses reasoning context with no signal. Anthropic's launch post names switching accounts mid-session in Claude Code as one case.2
GitHub made the model generally available in Copilot the same day, billed at provider list price, with a gradual rollout.3 Its own description is more modest than Anthropic's: it says Sonnet 5.5 matched Sonnet 5 on coding tasks while using significantly fewer steps, tokens and tool calls.3
Shipped
02GitHub Enterprise Cloud enforces the self-hosted runner minimum today, 29 September
GitHub moved the date forward in a changelog posted on 28 September.4 Runners below 2.329.0 cannot register or re-register. Runners below the higher minimum for running jobs stop picking up jobs, even when already registered. The change covers Enterprise Cloud only. Enterprise Server is unaffected, and Data Residency tenants have been enforced since 31 July.4
MCP TypeScript SDK 2.2.0 and 1.31.0 bind OAuth credentials to their issuer
Building the machine-to-machine OAuth providers without expectedIssuer is now deprecated and logs a warning. fetchToken() throws AuthorizationServerMismatchError when stored client information belongs to a different server.5 In 1.31.0, stored tokens gain an issuer field, so storage that rejects unknown fields must be changed to accept it.6 In 2.2.0, listTools() and the other list calls without a cursor now follow every page, capped by listMaxPages.5
Promised, not shipped
[object Object]
The conversation
01On Hacker News, the Sonnet 5.5 thread (704 points) mostly asked who needs a mid-tier model now.
- Sol-, Hacker Newsreader-posted, unverified
Said Opus 5.5 is efficient enough that the 5x plan covers two or three sessions at once, and asked when they would ever pick Sonnet 5.5.
- azuanrb, Hacker Newsreader-posted, unverified
Argued that below the frontier tier, GLM and DeepSeek do the job for a fraction of the price.
Anthropic's own table has Sonnet 5.5 at 46.2% on FrontierCode at Max effort and 52.1% at Xhigh, against Opus 5.5 at 54.4%.2 By list price Opus costs twice as much per token ($4/$20 against $2/$10).12 Whether Sonnet is cheaper per finished task depends on how many tokens each spends, and so far only the vendor has published that figure.
Hacker News was swept in full: 41 stories in the window, 12 above 80 points, read with their comments. Vendor changelogs and status pages: 12 feeds, 8 entries. GitHub releases across 48 watched repositories: 18 in the window. arXiv, 48 hours: 400 papers fetched, 169 kept, top 12 read. X did not run for this window. The newest keyword digest was collected 28 September 05:30 UTC and the newest roster digest (100 accounts) on 27 September 05:30 UTC, about 24.5 and 48.5 hours old, so both were background only. Reddit: the newest digest is also from 28 September 05:30 UTC, and the fallback pulse across six subreddits returned no posts.
Skip this
05Sonnet 5.5 beating Opus 5.5 on Terminal-Bench 4.0 Anthropic's number on Anthropic's run (70.6% against 66.4%), with a footnote on the Opus score. We quote their prices and leave the benchmark out.
Holo4 computer-use model Hugging Face post from the vendor, with no independent evaluation.
Nvidia's watchdog chip for agents Reported by CNBC. No part number, availability date or price we could read.
World Labs joining AMD; Anthropic IPO prospectus Corporate and financial news. Nothing a developer installs or pins changes.
Claude Code 2.1.284, Anthropic Python SDK 1.9.0, OpenAI Python 3.20.0 No breaking, deprecation or removal sections. Upgrade on your usual schedule.
Everything we saw
4141 candidates scanned · 5 used in this issue — the rest, with the reason each one was left out
| Item | Source | Signal | Call |
|---|---|---|---|
| Claude Sonnet 5.5 (Anthropic release notes, HN 704 points) | Vendor changelog | New model, five migration breaks | Led the issue |
| Claude Sonnet 5.5 in GitHub Copilot | GitHub Changelog | GA, gradual rollout | Used in the lead |
| Self-hosted runner version enforcement date has moved | GitHub Changelog | Enforcement 29 September | Printed as Breaking |
| modelcontextprotocol/typescript-sdk 2.2.0 and 1.31.0 | GitHub Releases | Upgrade notes: issuer binding, full-list walk | Printed as Breaking |
| Copilot Code Review unable to complete reviews | GitHub Status | Critical impact, resolved after 0.9 hours, 28 September 21:16 UTC | Resolved inside the window |
| Cursor status: Grok Bot degradation | Cursor Status | Minor impact, resolved after 0.2 hours | Weather |
| Cf: the agentic CLI for the Cloudflare API (HN 144 points) | Hacker News | New CLI launch | Primary page not opened this morning; checking tomorrow |
| arXiv 2609.35381: MCP error messages written for developers | arXiv | 949 of 3,001 error messages across 150 servers tell the reader to act outside the tools | Abstract only; worth a read for MCP server authors |
| Prompting Claude Opus 5.5 (HN 201 points) | Hacker News | Docs page, Opus 5.5 launched 22 September | Not new this window |
| vercel/ai 7.0.119, 7.0.122, 5.0.268; langchain 1.4.3, langchain-fireworks 1.7.0; crewAI 1.15.23; cline SDK 0.0.87 | GitHub Releases | Patch and minor releases, no breaking sections | No decision changes |
| BuilderIO/agent-native 0.1.272, clips 0.1.321, skills 0.3.10; stablyai/orca 1.4.216; alibaba/open-code-review 1.12.10 | GitHub Releases | Patch cadence | No decision changes |
| AI-lab accountability essays (Cal Newport, The Civilian, TechCrunch) and 'nobody knows the architecture' | Hacker News | 105 to 428 points | Opinion, no new measurement |
| OpenAI: Australia apology, Lenfest funding | OpenAI News | Corporate posts | Off the beat |
| X digests (keyword, roster) and Reddit digest | Community digests | Newest files 24.5 to 48.5 hours old | Background only; Reddit fallback returned nothing |
End of feed. That is everything from the window worth your time.
Explore more news, analysis and practical guides on NO FEED.