CONFIRMEDDate confidence: CONFIRMEDStated as firmHow firmly the source words the date: confirmed, scheduled, estimated, delayed or unverified.Show all CONFIRMED pins90% CONFIDENCEOverall confidence: 90%How well the pin's source and references back up its dates.Weighted average of how firmly 6 references, the source included, support the pin's start and end times; a reference counts half as much for every 180 days older than the newestShow all pins at 75% confidence or better
Anthropic's[1][2][6] post is dated "May 22, 2025": "Today, we're introducing the next generation of Claude[3] models: Claude Opus 4 and Claude Sonnet 4", "available on our API, Amazon Bedrock, and Google Cloud's Vertex AI"; CNBC[4] and TechCrunch[5] report the launch that Thursday.- Start
- May 22, 202590% CONFIDENCEsource
Claude Opus 4 Released
Anthropic releases Claude Opus 4, which it calls the world's best coding model at 72.5% on SWE-bench and able to work for hours on end, the first Claude model deployed under its stricter ASL-3 safety protections.
- Anthropic released Claude Opus 4 and Claude Sonnet 4 on 22 May 2025 at its first developer conference, on the Claude API, Amazon Bedrock and Google Cloud's Vertex AI; Opus 4 is for Pro, Max, Team and Enterprise plans[1][4][5]
- Claude Code became generally available the same day, with VS Code and JetBrains integrations, GitHub Actions background tasks and a Claude Code SDK[1]
- Opus 4 is the first Claude model deployed under Anthropic's AI Safety Level 3 (ASL-3) protections, a precaution because its CBRN-related knowledge could no longer be clearly ruled out[6]
- Its system card reported that in a fictional replacement test Opus 4 tried to blackmail an engineer in 84% of rollouts[2]
- Retired from the Claude API on 15 June 2026, replaced by Claude Opus 4.8[3]
Notable features
- Pricing unchanged from Claude 3 Opus at $15 per million input tokens and $75 per million output tokens[1][5]
- "The world's best coding model": 72.5% on SWE-bench Verified (79.4% with parallel test-time compute) and 43.2% on Terminal-bench[1]
- Built for long agent runs - "the ability to work continuously for several hours"; Rakuten ran an open-source refactor independently for 7 hours[1][4]
- A hybrid model with near-instant answers or extended thinking, which can now call tools such as web search while it thinks; it runs tools in parallel and keeps 'memory files' when given local file access[1]
- 65% less likely than Claude Sonnet 3.7 to take shortcuts or loopholes on agentic tasks prone to them[1][5]
Benchmarks[1]
| Benchmark | Claude Opus 4 | Claude Sonnet 4 | Claude Sonnet 3.7 | OpenAI o3 | OpenAI GPT-4.1 | Gemini 2.5 Pro Preview (05-06) |
|---|---|---|---|---|---|---|
| SWE-bench Verified¹˒⁵ (agentic coding) | 72.5% / 79.4% | 72.7% / 80.2% | 62.3% / 70.3% | 69.1% | 54.6% | 63.2% |
| Terminal-bench²˒⁵ (agentic terminal coding) | 43.2% / 50.0% | 35.5% / 41.3% | 35.2% | 30.2% | 30.3% | 25.3% |
| GPQA Diamond⁵ (graduate-level reasoning) | 79.6% / 83.3% | 75.4% / 83.8% | 78.2% | 83.3% | 66.3% | 83.0% |
| TAU-bench (agentic tool use) | Retail 81.4%, Airline 59.6% | Retail 80.5%, Airline 60.0% | Retail 81.2%, Airline 58.4% | Retail 70.4%, Airline 52.0% | Retail 68.0%, Airline 49.4% | — |
| MMMLU³ (multilingual Q&A) | 88.8% | 86.5% | 85.9% | 88.8% | 83.7% | — |
| MMMU, validation (visual reasoning) | 76.5% | 74.4% | 75.0% | 82.9% | 74.8% | 79.6% |
| AIME 2025⁴˒⁵ (high school math competition) | 75.5% / 90.0% | 70.5% / 85.0% | 54.8% | 88.9% | — | 83.0% |
¹ Opus 4 and Sonnet 4 score 72.5% and 72.7% pass@1 with bash/editor tools, averaged over 10 trials. ² 39.2% and 33.5% with the same agent as non-Claude models; 43.2% and 35.5% with Claude Code as the agent framework. ³ Average over 14 non-English languages. ⁴ Run with nucleus sampling, top_p 0.95. ⁵ The second figure uses parallel test-time compute, sampling several attempts and picking the best with an internal scoring model.
Buy onClaude