Copilot Chat: claude-sonnet-4.6 / claude-opus-4.6 max_prompt_tokens under-reported by CAPI — compaction fires 40K tokens early

[DarienSirius]

Summary

The GitHub Copilot API (/models endpoint, hereafter "CAPI") appears to report a max_prompt_tokens value for claude-sonnet-4.6 that is below the model's actual 200K context window as documented by Anthropic.

Impact

baseBudget in agentIntent.ts is set to Math.min(userSetting ?? endpoint.modelMaxPromptTokens, endpoint.modelMaxPromptTokens), where modelMaxPromptTokens flows directly from CAPI's capabilities.limits.max_prompt_tokens. If CAPI under-reports this value, the budgetThreshold = floor((baseBudget - toolTokens) * 0.85) fires early, triggering conversation compaction sooner than necessary — effectively wasting 40K tokens of usable context per request on every claude-sonnet-4.6 session.

Evidence

Per Anthropic's official model documentation, as of March 2026:

Model	Context window	Max output
claude-sonnet-4.6	200,000 tokens	64,000 tokens
claude-opus-4.6	200,000 tokens	128,000 tokens

The compaction threshold observable at runtime is significantly lower than 170,000 tokens (200K × 0.85), suggesting CAPI is returning approximately 160,000 for max_prompt_tokens.

Code path

CAPI /models → capabilities.limits.max_prompt_tokens
  → modelMetadataFetcher.ts _getMaxPromptTokensOverride()
    → chatEndpoint.ts this._maxTokens (line ~151)
      → registered as maxInputTokens with VS Code LM API
        → extChatEndpoint._maxTokens = languageModel.maxInputTokens
          → agentIntent.ts baseBudget
            → budgetThreshold = floor((baseBudget - toolTokens) * 0.85)

A user-side workaround exists (github.copilot.chat.advanced.summarizeAgentConversationHistoryThreshold) but the Math.min clamp in agentIntent.ts means it can only lower the threshold, not raise it above what CAPI reports. There is no client-side fix.

Expected behavior

CAPI returns max_prompt_tokens: 200000 (or ≥ 190000 accounting for the output token reservation logic) for claude-sonnet-4.6, and the compaction threshold fires at ~170K tokens.

Actual behavior

Compaction fires well below 170K tokens on claude-sonnet-4.6 sessions, consistent with a CAPI-reported max_prompt_tokens of ~160K instead of 200K.

Environment

• VS Code Insiders 1.110.0-insider
• github.copilot-chat 0.38.2026022001
• Model: claude-sonnet-4.6

Suggested fix

Update the CAPI model entry for claude-sonnet-4.6 (and claude-opus-4.6) to report max_prompt_tokens: 200000. No client code change required.

[bhavyaus]

We've made some updates to our compaction thresholds to help with this in updated releases.

[a11smiles]

This is also reported by the Enterprise API. Sonnet/Opus 4.6 allows 1M tokens. However, the API is capping requests to 168,000.

[bhavyaus]

Joshua Davis (@a11smiles) does this still repro on latest vscode insiders or 1.114 ?

[a11smiles]

Bhavya U (@bhavyaus) this is NOT using vscode. This is an error being generated by the GitHub Copilot Enterprise API. The error is one generated by the API endpoint.

[vs-code-engineering]

Thank you for creating this issue! Please provide one or more requestIds to help the platform team investigate. You can follow instructions found here to locate the requestId value.

Happy Coding!