By ByteBot
As of October 15, Anthropic moved Claude Opus 5.5 fast mode out of research preview. No waitlist, no access request, no special header required. If you have been waiting to use it or still carrying the old beta flag in production code, now is the time to clean that up.
What Actually Changed on October 15
Three concrete things shifted when fast mode hit GA:
- Waitlist gone. All Start, Build, and Scale tier API users can enable fast mode without requesting access.
- Beta header optional. Requests to Claude Opus 5.5 no longer need the
fast-mode-2026-02-01beta header. If your existing code still sends it, nothing breaks — it is just noise now. - Microsoft Foundry added. Fast mode landed on Claude in Microsoft Foundry for Azure deployments the same day. It still does not exist on Amazon Bedrock, Google Cloud, or AWS Platform.
One important distinction worth repeating: fast mode on Claude Opus 5.0 and 4.8 remains in research preview. Those models still require the beta header. Only Opus 5.5 graduated.
What Fast Mode Does (and Does Not Do)
Fast mode is not a different model. It runs the same weights as standard Opus 5.5 with a faster inference configuration. The speedup targets output throughput specifically — you get up to 2.5x more output tokens per second, which means a long streamed answer completes around 60% faster in wall time. What it does not change: time to first token, tool execution speed, or the adaptive thinking step. Opus 5.5 runs adaptive thinking continuously and that portion runs at normal speed regardless. Only the writing accelerates.
You can confirm which mode a response used by checking response.usage.speed, which returns either "fast" or "standard".
Migrating Your Code
The actual code change is minimal:
# Before GA: required beta wrapper and header
response = client.beta.messages.create(
model="claude-opus-5-5",
speed="fast",
betas=["fast-mode-2026-02-01"],
max_tokens=4096,
messages=[{"role": "user", "content": "..."}],
)
# After GA: standard client, no beta header
response = client.messages.create(
model="claude-opus-5-5",
speed="fast",
max_tokens=4096,
messages=[{"role": "user", "content": "..."}],
)
print(response.usage.speed) # "fast" or "standard"The old code still works. There is no forced migration, but cleaning it up avoids confusion for anyone joining your codebase later.
The 2x Price Tag: When It Is Worth It
Fast mode doubles the rate card — $8 per million input tokens and $40 per million output tokens, versus $4/$20 at standard speed. That math is simple. The use case question is less so.
| Standard | Fast Mode | |
|---|---|---|
| Input price | $4 / 1M tokens | $8 / 1M tokens |
| Output price | $20 / 1M tokens | $40 / 1M tokens |
| Output throughput | Baseline | Up to 2.5x |
| Time to first token | Same | Same |
| Amazon Bedrock | Yes | No |
Fast mode earns its premium in interactive sessions: live debugging, watching a large component generate, rapid prompt iteration where you are reading output as it arrives. It does not help in CI pipelines, batch jobs, tool-heavy agent turns where tests and builds dominate, or any workload you walk away from. For non-interactive work at scale, the Batch API at 50% of standard pricing is the sharper tool.
One gotcha: if you enable fast mode mid-conversation, the entire conversation context gets re-read at the uncached fast rate. Enable it at session start or not at all to avoid cost surprises in long agent runs.
Platform Availability
- Works: Claude API (all tiers), Claude Code, Claude in Microsoft Foundry
- Does not work: Amazon Bedrock, Google Cloud / Vertex AI, AWS Platform
- Claude Code specifics: Team and Enterprise plans require an org-level enable by an admin before individual users can activate it
Fast mode also runs on a separate rate limit pool. When capacity is exhausted, requests get a 429 error — not a silent fallback to standard speed. Build retry logic if you are building anything user-facing on top of it. The SDK retries up to twice automatically, but anything beyond that is your responsibility.
Bottom Line
If you are on the Claude API and your work involves long, streamed outputs that you read in real time, fast mode is now a straightforward toggle. Remove the beta header from Opus 5.5 requests, add speed="fast", and set that flag at session start. For everything else — batch work, tool-heavy pipelines, Bedrock deployments — standard mode still makes more economic sense.
