Anthropic ships Claude Opus 5.5.
claude-opus-5-5) — $4/$20, thinking always on
Claude Opus 5.5 lands at $4/$20 — thinking always on, multi-cloud, migration breaks
claude-opus-5-5 — Fable-class positioning at a lower Opus list price, same day on Claude API, Bedrock, Google Cloud, and Microsoft Foundry.List pricing: $4 / $20 per MTok (input/output); cache reads $0.20; 5m cache write $5 / 1h $8; Batch 50% off. Specs from platform overview: 1M context, 128k max output, adaptive thinking always on, default effort medium, training cutoff Jun 2026. Anthropic claims ~40% lower cost vs Opus 5 on typical workloads at default settings and >30% faster output — treat efficiency % as vendor.
Migration is real work: thinking cannot be disabled; tool_choice any/tool → 400; on Claude API and GCP, older computer tool → 400 (need computer_toolset_20260801; Bedrock still accepts the older tool). Cyber/bio safeguards track Fable 5.1 (many cyber tasks fall back to Opus 4.8). Flag announcement benches as vendor evals.
Why it matters: new default Opus SKU with price cut and breaking API changes on the same afternoon OpenAI cut Sol/Luna list — rewrite cost models and pin the migration checklist before flipping traffic.
Sources: announcement · platform overview · release notes · Foundry
None this window.
| Model / package | Org | Notes |
|---|---|---|
Claude Opus 5.5 (claude-opus-5-5)Sep 22 |
Anthropic | API $4/$20 · thinking always on · multi-cloud |
| GPT-6 Sol / Luna Sep 22 |
OpenAI | API $2/$10 · $0.10/$0.50 · Foundry GA |
| Switchyard v0.3.0 ~2:06 PM ET Sep 22 |
NVIDIA | GA Advisor Gate · breaking CLI removal |
| Ollama v0.34.4-rc0 ~10:24 PM ET Sep 22 |
Ollama | RC MLX Qwen 3.8 PP · Gemma 4 dyn res |
| llama.cpp tip b11126 ~8:01 AM ET Sep 23 |
ggml | Tip Vulkan IQ4_XS MMQ/MMV |
| Qwen3.6 DFlash2-4bit ~6:06 AM ET Sep 23 |
mlx-community | Hub community quant · not mlx GA |
| Item | Beat | Status |
|---|---|---|
| Opus 5.5 Fable-class cyber/bio Sep 22 |
Cyber often falls back to Opus 4.8 · Life Sciences / Cyber Verification paths | Watch |
| GCP CVP updated for Opus 5.5 Sep 22 |
Needs advancedAiEnabled + Anthropic data sharing |
Active |
| vLLM advisories ≤0.29.0 Sep 22 |
NIXL/Mooncake/disagg DoS-class — upgrade to 0.30.0 (Issue 22) | Upgrade |
| Claude inline-tools beta Sep 22 RN |
inline-tools-2026-09-15 mid-conversation tool add without busting cache |
Beta |
Critical · Opus 5.5 migration breaks
Thinking always on (disable → 400). tool_choice any/tool → 400. On Claude API/GCP, swap to computer_toolset_20260801 (Bedrock keeps older computer tool). Progress text between tools moves into thinking blocks — quiet UIs until display is adjusted.
Watch · GCP CVP Opus 5.5 · vLLM ≤0.29.0
GCP Cyber Verification for Opus 5.5 needs advanced AI and Anthropic data sharing. If you are still on vLLM ≤0.29.0 with NIXL/Mooncake/disagg, the Sep 22 advisories are why Issue 22’s 0.30.0 mattered.
Opportunity · Switchyard Advisor Gate · Genie MCP · GPT-6 cache
Smoke Switchyard 0.3.0 Advisor Gate + switchyard-server. Wire Genie One MCP for governed ontology context. Use GPT-6 prompt-caching pack (30-min prefixes, dashboard, effort/tool changes that preserve cache) on long agent loops.
Switchyard 0.3.0 Advisor Gate + switchyard-server
pip install 'nemo-switchyard==0.3.0' and cargo install --locked switchyard-server --version 0.3.0. Point a TOML deployment at executor + stronger advisor (Auto capable). Breaking vs 0.2 Python CLI.
Ollama 0.34.4-rc0 MLX Qwen 3.8 PP on Mac Studio
Install the RC; compare Qwen 3.8 MLX prompt-processing vs 0.34.3 GA. Optional: smoke mlx-community/Qwen3.6-35B-A3B-DFlash2-4bit under mlx-vlm 0.7.2 — community quant, verify arch support first.
Issue 23 · Wednesday, September 23, 2026 · Not investment advice · Sources cited inline · Other news: none this window · Issue 22 Grok 4.7 / MiMo / vLLM 0.30.0 / mlx-vlm 0.7.2 / Transformers GGUF / PyTorch 2.14 / b11101 not re-led
