[FEATURE] Automatic token/usage refund when a response fails to meet the request

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

When Claude or Claude Code produces a low-quality response (wrong code, ignored instructions, incomplete output), the tokens/usage quota consumed by that turn are still charged. Users pay twice: once for the failed attempt and again for the retry. There is currently no mechanism to reclaim usage for objectively failed responses.

Real-World Example:
Claude Design was asked to create a design system based on Bootstrap. Despite remaining quota being available, it stopped working after the Typography section. Additionally, the output was written entirely in English instead of the requested language. The result is unusable: the design tokens cannot be mapped to Bootstrap's functions/classes, and even within the delivered part (up to Typography), roughly half of the definitions are missing. The user paid full usage for a deliverable with zero practical value and must spend additional quota to redo the entire task.

Proposed Solution

Add a "refund tokens" function tied to the existing feedback mechanism (e.g., thumbs-down or a dedicated /refund command in Claude Code). When triggered, the tokens consumed by the flagged turn are immediately credited back to the user's quota. Optionally validate the claim automatically (e.g., detect ignored instructions, broken code, or contradiction with the prompt) to prevent abuse.

Alternative Solutions

None viable. Manual support tickets don't address the root cause. The goal of this feature is not just billing compensation — it is to create a direct financial feedback loop that incentivizes Anthropic to prevent low-quality responses in the first place. If bad turns cost Anthropic tokens instead of the user, quality failures become measurable and directly actionable in model/product development.

Priority

High - Significant impact on productivity

Feature Category

Other

Use Case Example

During long agentic sessions, a single bad turn (e.g., a design system generated incompletely and in the wrong language) consumes significant quota. I flag the turn, the quota is refunded, and I retry without paying twice.

Additional Context

Happened 4 times in one day with identical failure patterns.