Issue: Gemini 3.1 Pro charges tokens for safety-rejected requests

Gemini API
,
1

Observed Behavior

  • Gemini 3 Pro: Unsafe prompts are blocked at the API level
    (finishReason: SAFETY), no tokens consumed.
  • Gemini 3.1 Pro: Unsafe prompts are NOT blocked. Instead, a hidden
    safety prompt is injected into the context. The model runs full
    inference (including thinking) and generates a natural-language
    refusal. All tokens (input + injected safety prompt + thinking +
    output) are billed to the user.

Why This Is a Problem

  1. Cost: Users pay for inference that produces zero value.
  2. Transparency: The injected safety prompt is invisible to users.
    There is no metadata indicating the request was safety-filtered.
  3. Fairness: In the 3 Pro model, blocked requests cost nothing.
    The 3.1 Pro approach silently shifts safety infrastructure costs
    to the user’s token budget.

Suggested Fix

  • Option A: Restore hard-blocking for clearly unsafe prompts
    (no token consumption).
  • Option B: If soft-steering is preferred, add a response field like
    safety_prompt_injected: true and DO NOT bill tokens for
    safety-redirected responses.
  • Option C: At minimum, make the injected safety prompt visible in
    the API response metadata so users can audit their costs.
1 Like