Claude Opus 4 and Sonnet 4: What Anthropic Actually Shipped
Anthropic released Claude Opus 4 and Claude Sonnet 4 on May 22, 2025, aimed at coding and long-running agent work. Here are the published benchmarks, prices and features.
On this page
Correction (October 9, 2026): An earlier version of this article was dated before the models were released and contained benchmark scores, test results, customer figures and an anonymous quote that WikiDigit could not verify, including results attributed to "WikiDigit's own testing" that never took place. We have rewritten it using Anthropic's announcement, linked below.
Anthropic released Claude Opus 4 and Claude Sonnet 4 on May 22, 2025, presenting them as models built for coding, advanced reasoning and AI agents that work through long, multi-step tasks.
What are Claude Opus 4 and Sonnet 4?
They are the fourth generation of Anthropic's Claude models. Opus 4 is the larger, more expensive model meant for the hardest and longest tasks. Sonnet 4, which Anthropic called "a significant upgrade" to Sonnet 3.7, is the faster, cheaper option and, unlike Opus 4, was also available to free Claude users.
Benchmarks Anthropic published
All figures below are Anthropic's own, from its launch post:
| Benchmark | Opus 4 | Sonnet 4 | |---|---|---| | SWE-bench Verified (standard) | 72.5% | 72.7% | | SWE-bench Verified (high compute) | 79.4% | 80.2% | | Terminal-bench | 43.2% | not published |
SWE-bench Verified measures whether a model can resolve real GitHub issues in open-source Python projects. The "high compute" figures use parallel attempts with a selection step, so they are not directly comparable with the standard ones.
Price
| Model | Input (per million tokens) | Output (per million tokens) | |---|---|---| | Claude Opus 4 | $15 | $75 | | Claude Sonnet 4 | $3 | $15 |
Both were available through the Claude apps, the Anthropic API, Amazon Bedrock and Google Cloud's Vertex AI.
What was new besides the scores
- Extended thinking with tool use (beta): both models could alternate between reasoning and calling tools such as web search, instead of thinking first and acting afterwards.
- Parallel tool use, and in Sonnet 4's case, more precise instruction following.
- Memory files: when given access to local files, Opus 4 could write and update notes to keep track of long tasks.
- Fewer shortcuts: Anthropic said both models were 65% less likely than Sonnet 3.7 to use shortcuts or loopholes on agentic tasks prone to them.
- Claude Code became generally available, with integrations for VS Code, JetBrains and GitHub Actions.
Safety level
Anthropic said the release included "implementing measures for higher AI Safety Levels like ASL-3", a stricter tier of safeguards under its Responsible Scaling Policy.
What it meant
The launch made coding agents the center of Anthropic's pitch: the headline numbers were software-engineering benchmarks, and Claude Code shipped as a full product the same day. Benchmarks published by a model's maker are a starting point, not a verdict; independent evaluations and real use on your own codebase are a better guide.
More to Read
Unsealed Emails Show the Pentagon Told Anthropic Its Red Lines Were 'Just Not Workable'
Court emails show the Pentagon told Dario Amodei that Anthropic's guardrails on autonomous weapons were 'just not workable' — and that the blacklisting came the day after Amodei said the Pentagon's terms erased his red lines.
Claude Fable 5 Is Back Online. Its New Safety Filter Is Rerouting Debugging Work.
Anthropic's Fable 5 returned globally on July 1 after the export ban lifted, but a new jailbreak classifier cut its score on one third-party debugging benchmark by 70% almost overnight.
ChatGPT Just Lost Its Majority. The AI Race Is Finally On.
For the first time since its launch, ChatGPT's market share fell below 50%. Here's what the numbers say about where AI dominance actually stands.