In BYOK, usage reports 1.0 (should not do)

[lucasns97]

Summary

I'm using the Copilot SDK with BYOK (Anthropic provider) and receiving assistant.usage events with token usage data. The cost field appears to always be 1.0 regardless of actual token consumption, and I'd like clarification on what this field represents.

What I Observed

When monitoring assistant.usage events with claude-sonnet-4-5 via BYOK, the cost field consistently returns 1.0 regardless of token usage:

Test 1 - Short prompt (358 total tokens):

{
  type: "assistant.usage",
  data: {
    model: "claude-sonnet-4-5",
    inputTokens: 354,
    outputTokens: 4,
    cost: 1.0,  // Always 1.0
    duration: 1609
  }
}

Test 2 - Long prompt (813 total tokens):

{
  type: "assistant.usage",
  data: {
    model: "claude-sonnet-4-5",
    inputTokens: 382,
    outputTokens: 431,
    cost: 1.0,  // Still 1.0
    duration: 10110
  }
}

I also found issue #423 where gpt-5-mini shows cost: 0, suggesting the value varies by model.

Questions

1. What does the cost field represent?

   • Is it the actual dollar cost of the API call?
   • Is it a count of "premium requests" consumed (in GitHub Copilot's quota system)?
   • Is it a normalized cost unit for internal tracking?

2. For BYOK users calculating actual costs:

   • Should we rely on inputTokens and outputTokens for cost calculation?
   • Is the cost field meaningful when using BYOK, or does it only apply to standard GitHub Copilot authentication?

3. Why does it vary by model?

   • claude-sonnet-4-5: cost: 1.0
   • gpt-5-mini: cost: 0 (per issue gpt-5-mini not being recognized as valid model argument even though it works when set to default model in CLI #423)

Use Case

I'm building token usage tracking and cost analytics for multi-agent workflows using BYOK with Anthropic. Understanding what cost represents will help determine:

• Whether to use it in our cost calculations
• How to properly track expenses across different models
• Whether to implement our own cost calculation based on token counts

Environment

• SDK: Python (github-copilot-sdk)
• Authentication: BYOK with Anthropic
• Model: claude-sonnet-4-5
• Copilot CLI version: 0.0.403

Additional Context

The assistant.usage event provides excellent token tracking data (inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens, duration), which makes it very useful for usage analytics. Clarifying the cost field would complete the picture for developers building cost monitoring systems.

---

Question: What does the cost field represent?

[github-actions]

Thanks for opening this issue @lucasns97! This is a great question with excellent detail and examples.

I've labeled this as documentation and question since you're requesting clarification on the cost field in assistant.usage events. The issue applies to all SDKs, though I've added sdk/python since that's your environment.

The questions you've raised are important for developers building cost tracking systems, especially for BYOK users. A maintainer should be able to provide clarity on:

• What the cost field represents (dollar cost vs. quota units vs. normalized metric)
• Whether it's meaningful for BYOK scenarios
• Why it varies by model (1.0 for claude-sonnet-4-5, 0 for gpt-5-mini)

Your reference to issue #423 is helpful context. The team will follow up with documentation or clarification.

AI generated by Issue Triage Agent

[SteveSandersonMS]

Thanks for reporting this, @lucasns97. We agree this doesn't make sense in a BYOK scenario. I've retitled the issue to clarify we should not supply a value in this case.