BYOK - Anthropic Insane Costs

[andtii]

Im testing the SDK and tried using claude sonnet 4.6 with BYOK for a self improvement loop. After one hour it burnt 400$!!!! When using Github copilot plus with Copilot Cli and Opus 4.6 constantly in one month with extra credits is like 300$. What is BYOK doing different with the SDK vs Copilot CLI? Is there some obvious issues?

[github-actions]

Thank you for reporting this, @andtii! I've labeled this as a question and needs-info to help investigate your BYOK cost concerns.

To better understand what's happening, could you please provide:

1. Which SDK are you using? (.NET, Go, Node.js, or Python)
2. Session configuration details:
   • Are you using infinite sessions with persistence enabled?
   • What tools/functions do you have enabled?
   • Any custom system prompts or context being passed?
3. Self-improvement loop details:
   • How many iterations ran in that hour?
   • What's the typical prompt/context size per iteration?
4. **Token usage (redacted)
   • Can you share assistant.usage events showing inputTokens and outputTokens?
   • This will help us calculate whether the costs align with actual API usage

Why costs might differ from Copilot CLI:

• BYOK bills directly to your Anthropic account at their API rates
• Copilot Plus subscription uses GitHub's negotiated pricing and quota system
• Self-improvement loops can accumulate large context windows if session history isn't managed
• Infinite sessions may persist and grow context between iterations

Could you also check if your session is accumulating a large conversation history? You can monitor token counts via the usage events to see if context size is growing unexpectedly.

AI generated by Issue Triage Agent

[andtii]

Update: I dug into @github/copilot/index.js (core v0.0.420) and found why this happened — and a real SDK bug worth fixing.

Our fault: 74% duplicate dispatches

The main cost driver was our workflow bug — GitHub webhook events fired multiple times per issue, and we dispatched a new agent session each time. 42 tasks for 11 issues = 31 wasted sessions. We've fixed this with server-side dedup.

SDK bug: Anthropic BYOK response mapper drops cache token info

The SDK correctly sends cache_control: {type: "ephemeral"} on the system message and last tool — so prompt caching should be working at Anthropic's end.

But the response mapper ignores the cache fields from Anthropic's response:

// What the Anthropic BYOK mapper does (offset ~11149794 in index.js):
usage: {
  prompt_tokens: t.usage.input_tokens,
  completion_tokens: t.usage.output_tokens,
  total_tokens: t.usage.input_tokens + t.usage.output_tokens
}
// cache_read_input_tokens and cache_creation_input_tokens are never mapped

The OpenAI provider path (offset ~9338196) correctly maps t.usage?.input_tokens_details?.cached_tokens → prompt_tokens_details.cached_tokens. The Anthropic path doesn't do the equivalent.

Result: Every BYOK Anthropic user sees cacheReadTokens: 0 in all usage events, making it impossible to verify whether prompt caching is saving money.

This would be great to fix in the mapper — just map t.usage.cache_read_input_tokens and t.usage.cache_creation_input_tokens from the Anthropic response into the normalized usage format.

[SteveSandersonMS]

@andtii Thanks for looking into this, but I don't understand your conclusion. You're suggesting we need to fix the usage reporting, but that will have no impact on your actual API usage costs. Anthropic charges you based on what you actually use, not based on what our SDK reports to your app.

Did you get the $400 figure from actual Anthropic billing or were you estimating that based on the usage data that was returned to your app?

Or are you saying the $400 charge was actually correct due to your workflow bug, and that this SDK just needs to make a change to improve accuracy of usage reporting even though you know that won't affect how much you get billed?

[patniko]

Closing as part of a post-GA public backlog trim. This issue was identified as either already fixed/answered/superseded, or as stale low-priority backlog without current signal. If this still affects you on the current Copilot SDK/runtime, please open a new issue with updated reproduction details and we can re-triage it.