gpt-3.5-turbo-instruct, babbage-002, davinci-002, and gpt-3.5-turbo-1106 all stopped responding as of today and return errors. And a third consecutive day of papers on agent monitoring failure modes (CoT-monitor jailbreaking via transparent phrasing, not steganography) signals the research community is actively cataloguing the oversight infrastructure gap.gpt-3.5-turbo-instruct, babbage-002, davinci-002, gpt-3.5-turbo-1106 are gone; any code still calling them returns errors nowgit clone https://github.com/pacifio/atlas && cd atlas
bun install && cargo build --release
invalid_request_error 400s that didn't exist on Sonnet 5; one silent behavior change (text between tool calls moves into thinking blocks)."claude-sonnet-5" → "claude-sonnet-5-5" in your model ID without reading the migration guide will break any code that uses thinking: disabled, tool_choice: any/tool, computer use via computer_20251124, or an old-generation advisor pairing — 400 on the next request.thinking: {"type": "disabled"} → {"type": "between_tools"} — disabled returns a 400 on Sonnet 5.5. between_tools is the new minimum thinking level, available at high effort or below. At xhigh or max, use adaptive thinking (omit thinking or send {"type": "adaptive"}).tool_choice: {"type": "any"} or {"type": "tool"} → remove or use strict: true with auto — forced tool use returns a 400. Enforce tool calling via strict: true in strict tool use mode, or move the schema to structured outputs.computer_20251124 → computer_toolset_20260801 — on the Claude API and Google Cloud. On Bedrock, the old tool is still accepted.thinking blocks (empty at default display: "omitted"), not as text blocks. Applications streaming this text to users go silent between tool calls. Add thinking.display or switch to between_tools to restore it.Full before-and-after migration: https://platform.claude.com/docs/en/models/sonnet-5-5/migration-guide
claude-sonnet-5 (or plan to), use forced tool use, use thinking disabled mode, run computer use via the old toolset on the Claude API/GCP, or use the advisor tool with Opus 4.x advisors.EffortModerate — 6-point checklist; most codebases need 1-3 changes.gpt-3.5-turbo-instruct, babbage-002, davinci-002, and gpt-3.5-turbo-1106 — stopped responding as of today (September 28, 2026). API calls to these model IDs now return errors.gpt-5.6-terra (the published replacement for all four).completions endpoint models and the gpt-3.5-turbo-1106 snapshot. The replacement across all four is gpt-5.6-terra. Check your codebase for any direct model ID references (babbage-002, davinci-002 particularly appear in legacy fine-tuning, embedding, and text classification code). CI pipelines with model ID assertions will also fail if they still reference these names. These retirements were announced in advance and are not reversible.gpt-3.5-turbo-instruct, babbage-002, davinci-002, or gpt-3.5-turbo-1106.EffortQuick (for most: just swap model ID to gpt-5.6-terra). Moderate if you have fine-tuned models on the legacy base models that also need migration.high for agentic coding and hard tasks, medium for simpler ones. The 512-token minimum cache write (down from 1,024) means short prompt patterns that previously couldn't leverage caching now can.| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 70.6% | 10.3% | 66.4% |
| FrontierCode 1.1 | 46.2% | 42.4% | 54.4% |
| CursorBench 4.0 | 55.5% | 34.1% | 57.8% |
| GDPval-AA v2.1 | 1844 | 1449 | 1846 |
| Humanity's Last Exam | 64.5% | 54.9% | 67.7% |
| OSWorld 2.1 | 80.1% | 57.0% | 81.8% |
| SWE-Bench Pro | 81.3% | — | — |
Knowledge cutoff: June 2026. Context: 1M tokens. Max output: 128K tokens (300K on Batch API with beta header).
claude-sonnet-5-5 and the between_tools thinking type, promote cache diagnostics from beta to GA, and enable tools to run while the reply streams. TypeScript adds ClientToolUnion type for client-executed tools.between_tools thinking correctly to disabled on fallback hops.cache_diagnostics is now available on Message and MessageCreateParams in both SDKs without a beta header. If you've been monitoring cache hit rates for cost optimization, remove the beta header from your requests.between_tools fallback handling — if you use server-side fallback (fallbacks: "default") and a Sonnet 5.5 request falls back to Sonnet 5, the SDK now correctly degrades between_tools thinking back to disabled on the fallback hop. This matters because between_tools doesn't exist on Sonnet 5; without this fix, fallback requests would have errored.TypeScript additionally gains ClientToolUnion — a typed union for client-executed tools, useful when you have a mix of server-side and client-side tools in an agent loop.
pip install anthropic==1.9.0 / npm install @anthropic-ai/sdk@0.129.0. Cache diagnostics beta header removal is optional cleanup./usage command and status line now show dollar amounts alongside token counts; /mcp reconnect all command added; Google Cloud OTLP telemetry export added.claude-sonnet-5-5 unless you've pinned a model. The five breaking changes above now apply by default in Claude Code. If you've built automations or scripts that depend on Claude Code's Sonnet behavior (thinking modes, forced tool calls, computer use), verify them against the new model. (2) Dollar amounts in /usage and the status line — now shows e.g. "$271.40 / $500.00 spent this month" — makes budget tracking in gatewayed org deployments easier. (3) /mcp reconnect all is a useful new command for recovering from MCP server connection failures in interactive terminal sessions without restarting the session. (4) The auth: { google: {} } telemetry option exports gateway OTLP data to Google Cloud — relevant for teams running Claude Code in Google Cloud environments with OTLP-based observability stacks.@-mention system can pull prior session context into new sessions: @last-refactor-session injects the reasoning from a previous agent session as context for the current one. MCP integration means Claude Code natively discovers atlas context tools. Multi-agent parallel execution with shared memory (alpha feature) lets you run a fast Sonnet 5.5 session and a deeper Opus 5.5 session against the same repo simultaneously.No confirmed leaderboard movements in this scan period. lmarena.ai was egress-blocked; the September 2026 snapshot shows claude-opus-5-max at 1505 ELO and kimi-k3-max as the highest-rated open model at 1476 — neither change is specific to September 28. Claude Sonnet 5.5's benchmark data is covered in Model Releases above.
Note: Specific evasion rate numbers from arXiv:2609.31121 were not accessible in this scan (arxiv.org egress-blocked); entry scored at the Trends threshold (≥2) rather than the main digest threshold (≥3) for this reason.
Nothing cleared the quality bar this period. HN threads on the Sonnet 5.5 launch were not yet at score >200 at time of scan (model shipped ~4 hours before digest compilation). LessWrong posts on Monitor Jailbreaking (arXiv:2609.31121) were egress-blocked. simonwillison.net was egress-blocked.
curl -fsSL https://raw.githubusercontent.com/trailofbits/coop/main/install.sh | bash. github.com/trailofbits/coopAnnual developer conference at Fort Mason, San Francisco. Sam Altman keynote at 10am PT, livestreamed. No product announcements previewed. Note: Today's legacy model retirements (covered in Breaking Changes) appear to be a scheduled housekeeping action ahead of DevDay, not a DevDay announcement. [openai.com/index/devday-2026/]
Sonnet 5.5 launched today; Haiku 5.5 is the expected next follow-on. No announcement, no date, no benchmarks published. The gap between Opus 5.5 (September 22) and Sonnet 5.5 (September 28) was six days; Haiku's timeline is not confirmed. [inference from Sonnet 5.5 release pattern]
Migrate to antigravity-preview-09-2026, rename snake_case built-in tool parameters to PascalCase, and update file-edit handling. [ai.google.dev/gemini-api/docs/changelog]
Standard pricing begins; free billing period ends October 4. [platform.openai.com/docs/changelog]
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.