github-featured / comet-ml--opik

发布

0 发布 8,340 Git 标签

RSS 订阅源
  • 2.1.4-5813 8ecef16846
    比较

    [BE] fix: discount audio input tokens in cost for OpenAI audio models (#7137)

    frostbyte_neo 发布于 2026-06-25 14:56:48 +00:00

    • [BE] fix: discount audio input tokens in cost for OpenAI audio models

    23 OpenAI models in model_prices_and_context_window.json publish
    input_cost_per_audio_token at a rate distinct from input_cost_per_token
    (gpt-4o-audio-preview, gpt-4o-realtime, gpt-audio families). For
    gpt-4o-audio-preview the audio input rate is 16x the text input rate
    ($4e-5 vs $2.5e-6 per token). The Java backend had zero references to
    this JSON field, so audio input tokens were silently billed at the
    standard text input rate.

    Wires input_cost_per_audio_token through ModelCostData -> ModelPrice ->
    CostService, and updates SpanCostCalculator.textGenerationCost to read
    the audio token count from the OTel keys logged by the Python SDK
    (original_usage.prompt_tokens_details.audio_tokens with a bare-key
    fallback). When the configured audio rate is zero, behavior matches
    the previous code path; otherwise the (prompt_tokens - audio_tokens)
    remainder is billed at the standard input rate and the audio tokens at
    the configured audio rate.

    Test calculateCostBillsAudioPromptTokensAtAudioRateForOpenAiAudioModels
    mutation-verifies the new behavior: without the SpanCostCalculator
    change it fails (expected 0.01575, actual 0.0045). All 80 existing
    CostServiceTest and SpanCostCalculatorTest assertions remain green.

    Output audio tokens (output_cost_per_audio_token) are an intentional
    follow-up so this PR stays focused on the prompt side.

    Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com

    • [BE] test: cover audio-token branches with parameterized cases

    Address review feedback from #7137:

    • Move the audio-token narrative comment to a javadoc block on the test
      method (nit).
    • Refactor to @ParameterizedTest covering all three branches in the
      updated textGenerationCost:
      1. gpt-4o-audio-preview with original_usage.prompt_tokens_details.audio_tokens
        (Python SDK 1.6.0+ key) -> 0.01575
      2. gpt-4o-audio-preview with bare prompt_tokens_details.audio_tokens
        (OTel fallback key) -> 0.01575
      3. gpt-4o-mini (no input_cost_per_audio_token configured) so
        inputAudioRate == 0 and the audio key is ignored -> 0.00027

    Mutation-tested: reverting SpanCostCalculator to main makes both audio-rate
    cases fail (0.0045 vs 0.01575) while the no-audio case continues to pass,
    confirming each branch is independently exercised.

    Co-authored-by: Andres Cruz andresc@comet.com

    下载附件