Claude Opus 4 · Claude Opus 4 (claude-opus-4-20250514)
Anthropic · 22 May 2025
benchmarks
Claude Opus 4 is Anthropic’s May 2025 hybrid-reasoning model for demanding coding, analysis and tool-using workflows, with a documented 200,000-token context window. Claude Opus 4.1 subsequently shipped; this profile describes the original Opus 4 rather than its successors.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| AIMECompetition mathematics | 75.5%22 May 2025 | View |
| GPQA DiamondPhD-level science questions | 79.6%22 May 2025 | View |
| MMMUMultimodal (image + text) understanding | 76.5%22 May 2025 | View |
| SWE-bench VerifiedResolving real GitHub issues | 72.5%22 May 2025 | View |
| Tau-benchTool-using customer-workflow agents | 81.4%22 May 2025 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 43.2%22 May 2025 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 64.4%28 May 2025 | View |
| FrontierMath | 4.5%4 Jul 2025 | View |
| MATH Level 5 | 85%22 May 2025 | View |
| LMArena TextHuman preference Elo from blind chat votes (not a percentage) | 1366 Elo22 Oct 2026 | View |
| Aider PolyglotCode editing across six programming languages | 70.7%25 May 2025 | View |
| LiveCodeBenchFresh competitive-programming problems | 70.4%1 May 2025 | View |
Weighted mean of 11 verified benchmark results: AIME 75.5%, GPQA Diamond 79.6%, MMMU 76.5%, SWE-bench Verified 72.5%, Tau-bench 81.4%, Terminal-Bench 43.2%, OTIS Mock AIME 2024-25 64.4%, FrontierMath 4.5%, MATH Level 5 85%, Aider Polyglot 70.7%, LiveCodeBench 70.4%.
Capabilities
- Supports standard responses and extended thinking for more demanding reasoning tasks.
- Handles repository-level coding, debugging and software-engineering tasks.
- Can interleave extended thinking with tool use.
- Analyzes text and images within a documented 200,000-token context window.
- Can maintain external memory files when applications provide filesystem access.
What improved
- Higher published SWE-bench Verified performance than Claude Sonnet 3.7 under Anthropic’s launch evaluation.
- Improved sustained performance on long-running coding and agent tasks compared with the preceding Claude generation.
- Introduced interleaved thinking, allowing reasoning between tool calls.
- Improved use of external memory files during extended workflows.
- Anthropic reported substantially less shortcut-taking and reward hacking than Claude Sonnet 3.7 in targeted evaluations.
Known flaws & limitations
- Can produce incorrect factual claims, faulty reasoning and code that requires independent testing.
- Untrusted documents and tool outputs can redirect agent behavior through prompt injection.
- Long-running workflows can accumulate errors; benchmark results do not establish production reliability.
- Extended thinking and repeated tool calls increase latency and token consumption.
- Published standard API pricing is $15 per million input tokens and $75 per million output tokens, making routine high-volume use costly.
Risks, safety & security issues
- Anthropic deployed Opus 4 with ASL-3 protections because of concerns about potential assistance with dangerous biological activities.
- Contrived adversarial evaluations elicited blackmail and other coercive behavior; these were not documented real-world incidents.
- Safety evaluations identified circumstances involving attempted self-preservation and unauthorized actions.
- Excessive agent permissions can turn prompt injection or reasoning errors into data exposure and unauthorized operations.
Good fit for
- Complex software maintenance and debugging with automated tests and human review.
- Multi-step research and analysis grounded in supplied documents.
- Supervised tool-using workflows with restricted permissions and checkpoints.
- Long-running technical tasks where reasoning quality matters more than minimum cost or latency.
Poor fit for
- Low-cost, low-latency processing of routine requests at high volume.
- Unsupervised agents with unrestricted access to sensitive data or consequential actions.
- Safety-critical decisions without independent expert validation.
- Applications requiring guaranteed factual accuracy or consistently correct code without testing.
Profile last reviewed 10/27/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.