Anthropic is running an experiment: what if its flagship model just ran faster? The company quietly rolled out fast mode for Claude Opus 4.6, a research preview that pushes output token speeds up to 2.5x higher. No new weights, no new architecture. Just a faster inference configuration of the same model, available now in Claude Code and via an API waitlist.
The catch is the bill. Fast mode costs $30 per million input tokens and $150 per million output, a straight 6x multiplier over standard Opus 4.6 pricing ($5/$25). And that's only for prompts under 200K tokens. Cross that threshold and the multiplier doubles to 12x. Anthropic's own docs describe this as "a higher cost per token" with characteristic understatement.
There's a launch sweetener: 50% off through 11:59 PM PT on February 16. Even discounted, fast mode runs at triple standard Opus rates. Anthropic says its own engineers use the feature internally, which tracks. When you're building the model, you can expense the inference bill.
For Claude Code users, toggling /fast on requires extra usage enabled on your account. API access is more restricted, gated behind the waitlist with a beta header (fast-mode-2026-02-01). Fast mode doesn't work on batch API, priority tier, or third-party cloud providers like AWS Bedrock or Google Vertex. It also invalidates prompt cache, so switching between fast and standard mid-conversation gets expensive quickly.
GitHub Copilot is already rolling it out to Pro+ and Enterprise users. Cursor has it too.
The Bottom Line: Fast mode is Anthropic testing whether developers will pay 6x for speed on an already expensive model, and the answer will shape how they price inference tiers going forward.
QUICK FACTS
- Speed gain: up to 2.5x faster output tokens per second
- Pricing: $30/$150 per MTok (6x standard Opus 4.6 at $5/$25)
- Over 200K token context: 12x standard pricing
- 50% launch discount expires February 16, 2026 at 11:59 PM PT
- Available now in Claude Code (extra usage required), API via waitlist
- Not available on Bedrock, Vertex AI, or Azure Foundry
