What does fast mode change, and what does it leave alone?
Anthropic describes fast mode as the same model on a faster inference configuration, with no change to intelligence or capabilities (Anthropic, fast mode, read 19 November 2026). The gain is output tokens per second, not time to first token, so a long answer finishes sooner while the first word arrives about as fast as before. It shows most with streaming, where a person watches the text appear. For the first-word delay, see time to first token.
What does fast mode cost?
You pay a premium on every token, input and output, across the whole context window. The prices below are from the same page (read 19 November 2026), per million tokens.
| Model | Input | Output |
|---|---|---|
| Claude Opus 5.5 | $8 | $40 |
| Claude Fable 5.2 | $16 | $80 |
| Claude Mythos 5.2 | $16 | $80 |
Fast mode is not available with the Batch API or with a Priority Tier commitment. Falling back from fast to standard speed also misses the prompt cache, because the two speeds do not share cached prefixes. Speed is a purchase, so price it per finished task, not per token: see cost per task.
When is fast mode worth paying for?
Use it where a person is waiting on a long answer and the wait costs more than the premium. Examples are a coding assistant editing a large file, a live drafting tool, or an agent step on the critical path of a customer call. Leave it off for overnight batch work, background agents and short replies, where nobody sees the difference. Set the target first with a latency budget, then check whether output speed is the step eating it.
What does fast mode mean in Claude Code?
In Claude Code, fast mode is the same speed setting behind a toggle. Run /fast to switch it on or off, and Claude Code uses the same model at up to 2.5x the speed for a higher cost per token. On subscription plans it is billed through usage credits and is not included in the plan’s rate limits (Claude Code docs, fast mode, read 19 November 2026). The docs recommend it for interactive work such as live debugging, and standard speed for long autonomous tasks, batch jobs and cost-sensitive work.
What can go wrong with fast mode?
Fast mode has its own rate limit, separate from standard limits. Past it, the API returns a 429 error, and the SDK retries up to two times by default (Anthropic, read 19 November 2026). Decide in advance whether to wait or drop to standard speed, and watch the speed field in each response’s usage to see which one you got.
Availability moves too. The page lists fast mode as available on Opus 5.5, Fable 5.2 and Mythos 5.2, in research preview on Opus 5, and deprecated on Opus 4.8, where requests with speed: "fast" start returning an error after 30 November 2026 (read 19 November 2026). Any budget that assumes fast mode needs a date next to it.