Allow Max users to prefer model quality over speed when over capacity

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Summary:
When users exceed their weekly capacity limit, Claude automatically may route requests to smaller/faster models (e.g., Haiku instead of Opus). There is currently no user preference to opt for reduced speed while maintaining the same model quality.

Current Behavior:

• Hitting capacity limit triggers automatic model downgrade or queuing
• User has no control over this tradeoff
• No setting exists to express preference for quality vs. speed

Requested Behavior:

• Add a user preference: "Prefer quality over speed when over capacity"
• When enabled, keep requests on the user's selected model but accept longer queue times / slower responses
• Let users make the tradeoff decision themselves

Proposed Solution

Requested Behavior:

• Add a user preference: "Prefer quality over speed when over capacity"
• When enabled, keep requests on the user's selected model but accept longer queue times / slower responses
• Let users make the tradeoff decision themselves

Alternative Solutions

1. Settings Toggle: Add a preference in Settings → Subscription or Settings → Model Preferences:

• [ ] Prefer model quality over speed when over capacity
• Default: off (current behavior preserved)

1. Behavior When Enabled:

• Queue requests for the user-selected model instead of auto-downgrading
• Display estimated wait time so users can make informed decisions
• Allow per-request override (e.g., "Switch to Haiku for this request" option in UI)

1. Optional Enhancements:

• Show current queue position/wait estimate
• Allow setting a max wait threshold (e.g., "Wait up to 5 minutes, then downgrade")
• Per-conversation or per-request toggle for flexibility

Priority

Low - Nice to have

Feature Category

Performance and speed

Use Case Example

Users doing complex development work (code review, architecture, debugging) may prefer waiting longer for Opus-quality responses rather than getting faster but less capable responses from a smaller model.

Additional Context

_No response_