Claude Opus 4.6 review: when the flagship beats Sonnet 4.6
Claude Opus 4.6 (anthropic-claude-opus-4-6) is the model for hard tasks where quality comes first. It also arrived (February 5, 2026) with the smallest gap to its sibling Sonnet 4.6 that Anthropic has shipped. The sensible setup: route the hard cases to Opus and leave everyday work to Sonnet or Haiku.
Anthropic has released newer Opus models since (Claude Opus 5, then Opus 5.5), and FastMetal serves them. Opus 4.6 is still served.
What's new in 4.6
Opus is the top tier of the Claude family, built for complex reasoning, long documents and careful coding. Anthropic's 4.6 release adds:
- Adaptive thinking: four effort levels (low, medium, high, max; high is the default), and Claude scales its reasoning to the difficulty of the task. Anthropic says low cuts output tokens by about 40%, which saves money on light work.
- Up to 1 million tokens of context in beta, and up to 128K output tokens.
- Agent Teams: several Claude instances working in parallel. Opus only.
Strengths
- Agentic coding: Anthropic reports 80.8% on SWE-bench Verified, 65.4% on Terminal-Bench 2.0 and 72.7% on OSWorld (computer use), each a new high for the company at release.
- Multilingual coding: 77.83% on SWE-bench Multilingual.
- Very long inputs: with the 1M context beta it can take a large codebase or a stack of documents in one request.
Weak spots
- Cost: Opus costs much more per token than Sonnet, so running every request through it adds up fast. Compare the Opus 4.6 and Sonnet 4.6 pages.
- A narrow lead: on the main benchmarks Sonnet 4.6 is close behind, so for most work there is little reason to pay for Opus.
- No per-language scores: Anthropic publishes no Japanese-only (or other single-language) benchmark.
- Where it runs: FastMetal serves Opus 4.6 from a global region, while Sonnet 4.6 and Haiku 4.5 are served in Japan. If data location matters to you, the model page says which.
Opus 4.6 vs Sonnet 4.6
Reviewers put the SWE-bench Verified gap at 1.2 points, the smallest in Claude's history. Sonnet 4.6 is reported to deliver 97 to 99% of Opus 4.6's coding performance at a far lower cost.
| Opus 4.6 | Sonnet 4.6 | |
|---|---|---|
| SWE-bench Verified | 80.8% | 79.6% |
| Relative cost | Higher (top tier) | Lower (mid tier) |
| Agent Teams | Yes (Opus only) | No |
| Best for | High-stakes changes, large migrations, security review | Everyday fixes, new features, routine refactors |
Figures are from Anthropic and published reviews.
Cost
Opus is billed per million input and output tokens at a top-tier rate, and the spread to Sonnet stays wide, so splitting work by difficulty is the starting assumption. Current rates are on the pricing page. FastMetal is prepaid: you are only charged up to the balance you loaded.
Outside English
In our use it reads and writes Japanese well and holds up on involved instructions and long context. We can't give a Japanese score, though, because Anthropic doesn't publish one. For non-English work, start with Sonnet 4.6 and move only the cases it struggles with up to Opus.
Where it fits
Good fit: high-risk refactors, large code migrations, security review, analysis and decision support grounded in long documents.
Overkill: summarizing, classifying and other light tasks (use Haiku 4.5), and the everyday balance of fixes and features (use Sonnet 4.6).
FAQ
Should I use Opus 4.6 or Sonnet 4.6?
For most work, Sonnet 4.6. The SWE-bench gap is 1.2 points, and Sonnet is reported to deliver 97 to 99% of Opus's performance at a lower cost. Move up to Opus for the cases where a mistake is expensive: high-stakes changes and large migrations.
Can anyone use the 1M-token context?
Anthropic offers the 1M context in beta, with up to 128K output tokens. It helps when you need a whole codebase or a long document set in one request, and the cost grows with the context, so send only what the task needs.
What is adaptive thinking?
Claude adjusts how much it reasons to the difficulty of the task. There are four levels (low, medium, high, max), high is the default, and Anthropic says low cuts output tokens by about 40%.
Try it on FastMetal
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "anthropic-claude-opus-4-6",
"messages": [{"role": "user", "content": "Propose three designs that meet this spec: ..."}]
}'
Opus for the hard cases, Sonnet or Haiku for the rest: one API key switches between them. Every model is in the catalog; rates are on the pricing page.