Support fast mode in headless (-p) mode
Summary
Fast mode (/fast) provides ~2x throughput improvement in interactive sessions but has no effect in headless (-p) mode. There's no CLI flag to enable it for programmatic use.
Evidence
Benchmarked 500-word essay generation with Opus 4.6, 3 runs each, alternating:
| Mode | Avg tok/s |
|------|-----------|
| Interactive (fast OFF) | ~42 tok/s |
| Interactive (fast ON) | ~2x faster (wall time: 14.8s vs 6.9s) |
| Headless fastMode: true in settings | ~42 tok/s (no change) |
| Headless --betas fast-mode-2026-02-01 | ~41 tok/s (no change) |
The API requires both anthropic-beta: fast-mode-2026-02-01 header AND speed: "fast" in the request body. The CLI's --betas flag can set the header, but there's no --speed flag to set the request body parameter.
Request
Add one of:
--fastflag forclaude -pmode--speed fastCLI parameter- Have
fastMode: truein settings actually apply to headless mode
Use Case
Autonomous agent systems (like CI/CD pipelines or orchestration tools) that spawn claude -p workers would benefit significantly from 2x throughput. Currently the only workaround is bypassing claude -p entirely and calling the Anthropic API directly, which loses the agent tool loop.
Environment
- Claude Code (latest)
- macOS, Opus 4.6
- Using
claude -p --model opus --output-format json
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗