Claude 4.6 appears capped at 200k input context while direct Anthropic supports 1M #512
Closed as not planned
Description
What's happening
Claude 4.6 through opencode-antigravity-auth behaves like a ~200k input lane. Once input history gets near that point, sessions compact or fail with prompt-too-long style errors.
The same model family on direct Anthropic can go beyond 200k when long-context beta is enabled (context-1m-2025-08-07) and the account is eligible.
Why this is confusing
From a user perspective, the model name is the same, so people expect similar context behavior. In practice, effective limits differ by provider path.
Requested changes
- Confirm whether the antigravity path can expose/pass the equivalent long-context capability for supported Claude models.
- If it cannot, make the 200k cap explicit in provider model metadata and UI text.
- Show provider-path-specific context limits in model selection so users can choose intentionally.
Repro (minimal)
- Use Claude 4.6 via antigravity-auth.
- Keep growing input context past ~200k.
- Observe compaction/overflow behavior.
- Compare with direct Anthropic + long-context beta enabled.
Current workaround that works
Until this is fixed in provider behavior/docs, this split setup is reliable:
- Keep
google/antigravity-*as the 200k lane. - Use
anthropic/*for long-context runs. - Inject
context-1m-2025-08-07only for direct Anthropic requests via plugin hooks. - For direct Anthropic model overrides, set both
limit.context=1048576andlimit.input=1048576.
High-level plugin logic:
- If
input.model.providerID !== "anthropic", do nothing. - For Claude 4.x long-context-capable models, add
anthropicBetaplusanthropic-betaheader valuecontext-1m-2025-08-07.
This keeps antigravity behavior unchanged while enabling a 1M path for direct Anthropic where entitlement allows it.