GitHub Copilot Claude model limits are inaccurate — missing max_prompt_tokens and wrong context values

[maximharizanov]

Problem

The github-copilot provider entries for Claude models have inaccurate limit values. The data currently reports a single context value that doesn't reflect the actual limits exposed by the Copilot API (CAPI).

Source of truth

The authoritative source for Copilot model limits is the CAPI endpoint GET https://api.githubcopilot.com/models. The response includes three distinct token limits per model:

"family": "claude-opus-4.6",
"limits": {
    "max_context_window_tokens": 144000,
    "max_output_tokens": 64000,
    "max_prompt_tokens": 128000,
    "max_non_streaming_output_tokens": 16000
}

Current models.dev values vs CAPI (Claude models)

Model	models.dev context	models.dev output	CAPI max_context_window_tokens	CAPI max_prompt_tokens	CAPI max_output_tokens
claude-opus-4.6	128,000	64,000	144,000	128,000	64,000
claude-sonnet-4.5	128,000	16,000	144,000	128,000	32,000
claude-opus-4.5	128,000	16,000	160,000	128,000	32,000
claude-sonnet-4	128,000	16,000	216,000	128,000	16,000
claude-haiku-4.5	?	?	144,000	128,000	32,000

Every Claude model has discrepancies. The pattern is:

1. context is underreported — models.dev uses 128k, but CAPI reports 144k–216k for max_context_window_tokens
2. output is underreported for sonnet-4.5 and opus-4.5 (16k vs 32k actual)
3. max_prompt_tokens (input limit) is missing entirely — CAPI provides a separate max_prompt_tokens field (128k for all Claude models), which is different from max_context_window_tokens

Impact

This data is consumed by OpenCode (via https://models.dev/api.json) to determine compaction thresholds. The compaction logic in OpenCode uses:

const usable = model.limit.input 
    ? model.limit.input - reserved 
    : context - reserved

Since limit.input is not set (because models.dev doesn't populate it), OpenCode falls back to context - reserved, using the underreported 128k value. This triggers compaction earlier than necessary, wasting ~16k tokens of usable context on every session.

Suggested fix

For github-copilot Claude models, update the limits to match CAPI. The models.dev schema already supports limit.input — it just needs to be populated:

"claude-opus-4.6": {
    "limit": {
        "context": 144000,
        "input": 128000,
        "output": 64000
    }
}

How to verify

Query the CAPI models endpoint with a valid Copilot token:

curl -s "https://api.githubcopilot.com/models" \
  -H "Authorization: Bearer $COPILOT_TOKEN" \
  -H "Copilot-Integration-Id: vscode-chat"

This returns the full model catalog with authoritative limits per model.

[rekram1-node]

I'm not really sure why we should even track the max_context_window_tokens for all intensive purposes max_prompt_tokens is the real limit, afaik even github copilot treats it as the window for compaction logic

[maximharizanov]

I think it warrants some testing as GH Copilot isn't terribly helpful with what these terms mean. E.g. If my session is at 128k context, can I still initiate compaction (because I still have 16k output tokens remaining until I hit the max_context_window_tokens)? It would make a significant difference with Opus on Github Copilot as it's already at a tiny context window size.