Anthropic ships Claude Sonnet 5.5.
Claude Sonnet 5.5
claude-sonnet-5-5) — the second Claude 5.5 family model after Opus 5.5. List price stays $2 / $10 per MTok (same as Sonnet 5); Anthropic claims up to 30% lower cost per task and 30%+ faster generation. Same-day multi-cloud: Bedrock, Claude Platform on AWS, and GovCloud.Specs: 1M context, 128K max output, adaptive thinking (default effort high). First Sonnet with Opus-class cyber safeguards — higher-risk cyber tasks visibly fall back to Sonnet 5; cyberdefenders can apply to an expanded Cyber Verification Program. Migration note: to turn off up-front thinking send thinking: {"type": "between_tools"} (not "disabled"). Haiku 5.5 “coming weeks” — not launched.
Vendor benches (Terminal-Bench 4.0, CursorBench, GDPval-AA) are vendor / third-party eval mix — prefer “Anthropic reports…” wording. Business folded into this lead (same sticker price, claimed lower invoice). Do not re-lead Opus 5.5.
Why it matters: a mid-tier Sonnet SKU that ships cyber-tier safeguards and multi-cloud on day one — migrate carefully, watch token efficiency claims on your own workloads.
Sources: Anthropic · Platform RN · AWS WN · AWS ML blog
None this window.
| Model / package | Org | Notes |
|---|---|---|
| GPT-6.1 Sol Sep 29 · DevDay · gpt-6.1-sol |
OpenAI | New $2/$10 · $0.10 cached · Work/Codex/API · not Chat · vendor benches |
| Claude Sonnet 5.5 Sep 28 · multi-cloud same day |
Anthropic | New $2/$10 · 1M ctx · cyber safeguards · vendor benches |
| Ollama v0.35.0 Sep 28 · pre-release |
Ollama | Pre Decision Models · /v1/systemone · Nimble/Tev1 |
| MLX v0.32.3 Sep 28 evening ET |
Apple ML | GA GDN Metal · M5 Ultra · SDPA D512 |
| mlx-vlm 0.7.4 Sep 28 |
Blaizzy | GA DeepSeek V4.1 vision · InternVL3.5 |
| llama.cpp tip b11240 Sep 28 · multimodal embeddings |
ggml | Tip typed content on /v1/embeddings · GA still v0.5.0 |
| Grok 4.7 on Bedrock Sep 28 |
AWS / xAI | Cloud us. / global.xai.grok-4-7 · availability only |
| Decisions API Sep 29 · DevDay · limited preview |
OpenAI | Preview Luna-focused finite-choice answers · broad release “coming days” |
| Agents API + computer use Sep 29 · DevDay |
OpenAI | New Computer use + Codex multi-agent / tool search / compaction in API |
| Item | Beat | Status |
|---|---|---|
| OpenAI safety cases (frontier RL) Sep 28 3:00 PM ET |
Draft guidelines · vetoes · fail-closed · not regulation | Watch |
| Sonnet 5.5 cyber / CVP Sep 28 |
First Sonnet with Opus-class cyber fallbacks · expanded CVP | Watch |
| GPT-6.1 Sol safety addendum Sep 29 |
Critical cyber · High bio/chem · same safeguards stack as Astra | Watch |
| OpenAI DevDay / GPT-6.1 Sol Sep 29 |
Sol $2/$10 · Critical cyber addendum · dots Pro/Biz Premium · 20+ Recap items | Shipped |
| Jev calibration ongoing |
Typed decisions · confidence ≠ accuracy on adversarial pairs | Watch |
Critical · Sonnet 5 migrate (between_tools) · safety-case veto culture
Sonnet 5 → 5.5: send thinking: {"type": "between_tools"} (not "disabled"); forced tool_choice any/tool can 400. OpenAI’s safety-case note asks for senior vetoes and fail-closed monitors on frontier RL — treat as lab practice, not a pause lift.
Watch · Sol vendor benches · Jev calibration
GPT-6.1 Sol pricing is clear; near-Astra coding/computer-use scores are OpenAI-reported — smoke your own evals. Jev / Ollama systemone: typed outputs reduce format bugs; they do not prove calibrated confidence. Keep Ollama 0.35.0 marked pre-release.
Opportunity · Ollama systemone · Transformers GGUF on Mac
ollama pull nimble and hit /v1/systemone for a typed triage smoke. On Apple Silicon, try Transformers GGUF load (Python path) or mlx-vlm 0.7.4 DeepSeek V4.1 vision — llama.cpp stays the dedicated engine pin.
Ollama Decision Models — pull nimble + /v1/systemone
Install Ollama v0.35.0 (pre-release). ollama pull nimble, then call /v1/systemone with a choice or score question. Confirm typed choices/probabilities — not chat prose. Keep a GA pin on 0.34.4 for production until non-pre 0.35 ships.
Migrate Sonnet 5 → 5.5 (between_tools) or mlx-vlm 0.7.4 vision
Point a staging client at claude-sonnet-5-5; send thinking: {"type": "between_tools"} if you previously disabled thinking. Or on Apple Silicon: bump mlx-vlm to 0.7.4 and smoke DeepSeek V4.1 vision. Flag Anthropic efficiency claims as vendor until you measure your own tasks.
Issue 27 · Tuesday, September 29, 2026 · Not investment advice · Sources cited inline · Other news: none this window · Issue 26 OpenShell / Completions / pause / b11228 / Holo4 not re-led · Issue 25 Ollama 0.40.0-rc0 / mlx-vlm 0.7.3 / Live Avatar / AU Medicare not re-led · Issue 24 llama.cpp v0.5.0 / Ollama 0.34.4 / vLLM Sep 23 GHSAs not re-led · Issue 23 Opus 5.5 / Sol·Luna launch / Grok 4.7 launch / Switchyard not re-led · DevDay patch applied (Sol + Recap + dots) · Do not invent Haiku 5.5 GA · Listen omitted (episode not published)
Past 1–2 days. ASTGL / Fraser takes — sourced, not recap theater.
Sonnet 5.5 — same sticker, different invoice
Anthropic kept the $2/$10 list and is selling efficiency: fewer tokens per task, 30%+ faster output, plus the first Sonnet with Opus-class cyber fallbacks. That is a different invoice only if your workloads actually shrink — measure yourself. Migrate carefully:
thinking: {"type": "between_tools"}, not"disabled". Haiku 5.5 is still vapor on the launch page.Sources: Anthropic · Platform RN
Ollama Decision Models / System One locally
v0.35.0 (pre-release) adds
/v1/systemone— typed choices, scores, and probabilities via Nimble / Tev1, not free-form chat. This is the local mirror of TypeSafe’s Jev shape. Pullnimble, hit the decision endpoint, treat calibration as unproven until you measure it.Source: Ollama v0.35.0
OpenAI safety cases for frontier RL
After last week’s pause-and-disclose pack, OpenAI posted early guidelines for structured safety cases before continuing frontier RL runs: alignment training, containment, monitoring, plus senior vetoes and fail-closed ops. Guidelines, not regulation — and they do not say the tool-use pause ended.
Source: OpenAI safety cases
MLX stack wakes
mlx 0.32.3 lands Metal Gated DeltaNet kernels and M5 Ultra matmul tunings; mlx-vlm 0.7.4 adds DeepSeek V4.1 vision (and InternVL3.5). After a quiet stretch, Apple Silicon operators finally have a same-window core + VLM bump worth smoking.
Sources: MLX v0.32.3 · mlx-vlm 0.7.4
DevDay ships — Sol is the operator SKU
OpenAI’s DevDay recap lists 20+ announcements; the Monday move for API teams is GPT-6.1 Sol at $2 / $10 per MTok (cached input $0.10) — one-fifth of Astra’s standard $10/$50 — via
gpt-6.1-sol, ChatGPT Work, and Codex (not Chat yet). OpenAI’s own benches put Sol near Astra on coding / computer use / professional work; treat those as vendor. Safety addendum: Critical cyber / High bio-chem, same stack as Astra. Dots (Astra-powered always-on agents) roll to Pro / Business Premium in eligible markets — interesting for long-running work, not a local-runtime story.Sources: GPT-6.1 Sol · DevDay Recap · Safety addendum · dots