Support Anthropic fast mode (priority service tier)

[charles-cooper]

Summary

Claude Code introduced "fast mode" — same Opus 4.6 model but with an API configuration that prioritizes speed over cost efficiency. It uses the service_tier parameter on the Anthropic Messages API.

Reference: https://code.claude.com/docs/en/fast-mode

Current SDK state

• @anthropic-ai/sdk 0.74.0 (latest) supports service_tier?: "auto" | "standard_only" on request, and "standard" | "priority" | "batch" on response.
• service_tier: "auto" opts into priority routing when available.
• pi-mono currently pins SDK 0.71.2, package.json specifies ^0.73.0.

What to implement

packages/ai

1. Upgrade @anthropic-ai/sdk to latest (0.74.0+) to get current service_tier types.
2. Add serviceTier option to AnthropicOptions — pass through to the API as service_tier.
3. Expose in SimpleStreamOptions so callers can opt in without using provider-specific options.
4. Surface service_tier from response in usage/metadata events so callers know which tier was actually used.

packages/coding-agent

5. Add /fast slash command (or equivalent toggle) that sets service_tier: "auto" on the Anthropic provider.
6. Persist fast mode across sessions (settings file).
7. Show indicator in status line when fast mode is active.
8. Auto-fallback: if rate-limited on priority tier, fall back to standard and indicate cooldown.

Pricing context (for docs/UI)

Mode	Input (MTok)	Output (MTok)
Fast (<200K)	$30	$150
Fast (>200K)	$60	$225

Fast mode is billed to extra usage even if plan usage remains. Only available via Anthropic API (not Bedrock/Vertex/Azure).

Notes

• The SDK input type may expand beyond "auto" | "standard_only" in future versions. Watch for updates.
• Fast mode can be combined with effort level for maximum speed on simple tasks.
• Consider whether this should generalize to other providers that have similar tiering (e.g., OpenAI service_tier).

[badlogic]

That's a lot of plumbing for something I won't use, and it doesn't transfer to other providers cleanly.

You can create a custom provider that implements this. There's even a full anthropic example including streaming that can do this.

[charles-cooper]

i think the whole /fast machinery is optional, but being able to inject api headers (and the concept of 'priority' does exist for e.g. openai - https://platform.openai.com/docs/pricing) is generally useful i think

[leofisG]

Will be great if this feature is implemented for both Anthropic and OpenAI