Problem

Claude Opus 4.6 supports a 1M token context window, but the Anthropic API requires the beta header context-1m-2025-08-07 to unlock it. Without this header, the API enforces a hard 200k input token limit.

OpenCode currently sends this beta header for claude-sonnet-* models but does not include claude-opus-4-6. This causes the API to reject requests exceeding 200k tokens:

prompt is too long: 201965 tokens > 200000 maximum

Root Cause

The beta header logic only matches Sonnet model IDs. Opus 4.6 is the first Opus-class model with 1M context support and needs to be added to the header inclusion check.

Impact

• Config overrides setting limit.context: 1000000 have no effect — the limit is enforced server-side by the Anthropic API
• The status bar shows usage against 1M (e.g., 11% at 109k tokens), but the API rejects at 200k
• Auto compaction doesn't trigger in time because OpenCode believes the model has 1M of headroom

Expected Behavior

OpenCode should send anthropic-beta: context-1m-2025-08-07 header for claude-opus-4-6 requests, matching the behavior already in place for Sonnet models.