Claude Opus 4.7
Legacy (still served)
Superseded by Claude Opus 5.5. Status per Anthropic docs on 2026-09-30: active/legacy, retirement not sooner than 2027-04-16.
by Anthropic. Released 2026-04-16.
| API ID | claude-opus-4-7 (Bedrock anthropic.claude-opus-4-7) |
| Context / output | 1M tokens / 128K (300K on Batches API with beta header) |
| Thinking | adaptive (default effort high) |
| Knowledge cutoff | Jan 2026 |
| Platforms | Claude API, Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS |
Introduced the current tokenizer (1M tokens is about 555k words); from 4.7 on, non-default temperature/top_p/top_k are rejected by the API. Successor: Opus 4.8.
Earlier: Opus 4.6. See anthropic-claude for the full lineup.
Merged from laptop copy (2026-10-01)
Video-coverage notes from the owner’s earlier version of this note (laptop copy; claims come from channel summaries, not vendor docs, so treat as unverified):
| Date | Event |
|---|---|
| 2026-03-31 | Leaked alongside Sonnet 4.8, with “undercover mode” and hidden-model rumours |
| 2026-04 to 2026-06 | Active benchmark reference |
| 2026-06 | Superseded by Opus 4.8 |
Opus 4.7 had a roughly three-month window as Anthropic’s frontier reference (2026-03-31 to 2026-06-12 in source coverage) before 4.8 took over. Like its successor, most of its coverage is challengers claiming to beat it: MiniMax M3 (“50x cheaper”), DeepSeek V4, Nex-N2, Qwen 3.7 Max.
It also appears in the “agents refuse human control” discussion alongside GPT-5.5, an early instance of the autonomy-safety framing that grew through 2026.
Part of the Claude Opus 4 series.
Sources
- https://platform.claude.com/docs/en/models/opus-4-7/overview (accessed 2026-09-30)
- https://www.anthropic.com/news/claude-opus-4-7 (accessed 2026-09-30, linked from docs; not fetched)
- https://platform.claude.com/docs/en/about-claude/models/overview (accessed 2026-09-30)
- https://platform.claude.com/docs/en/about-claude/model-deprecations (accessed 2026-09-30)