Claude Opus 5.5 fast mode exits preview: 2.5x speed, double the price
Preview's over; it's live now, with real caveats on clouds and first-token latency. Plus Moondream 3.1, a free GLM playground, and DeepSeek V4.1 Flash.
Copy markdown
Fast mode is GA — same weights, 2.5x the throughput
Anthropic dropped the waitlist on Oct 15, ending the research preview. You get up to 2.5x output throughput on the same Opus 5.5 weights for $8/M input and $40/M output — exactly double standard Opus 5.5's $4/$20. It's a speed dial, not a smarter model.
Where you can — and can't — turn it on
Live in Claude Code and on the Claude Platform/API, including Managed Agents and Microsoft Foundry — but not Amazon Bedrock, Claude on AWS, or Google Cloud. Fast-mode rate limits are tracked separately from standard Opus, so a 429 there won't eat your normal quota.
The catch: your first token is just as slow
The speedup only kicks in after generation starts — time-to-first-token is unchanged. Switching between fast and standard busts your prompt cache, and the fallback-on-429 is something your app has to catch and retry itself. Great for long streamed outputs, no help for snappier latency.
Moondream 3.1: a 9B open VLM for your own box
A 9B-total / 2B-active MoE vision model with native query, detect, point and caption, all returning structured output. Runs locally on Ampere-or-newer NVIDIA GPUs and Apple Silicon via `pip install moondream`, no API key for the base model. Weekend build: an on-device screenshot-to-JSON pipeline.
Z.ai opens a free, no-key GLM playground
Z.ai put its MIT-licensed GLM models — Base, Reasoning and Rumination — behind a free official playground with no local setup or API key. The fastest way to check whether an open model can stand in for a paid one in your stack before you wire anything up.