Anthropic's Claude Opus 4.7 Retakes the Lead on Coding Benchmarks

Anthropic shipped Claude Opus 4.7 on April 16, and the headline result is straightforward: it's the most capable publicly available language model right now, at least on the benchmarks that matter most for coding and long-running agent tasks.

The numbers are hard to argue with. Opus 4.7 scored 87.6% on SWE-bench Verified, which tests real-world software engineering tasks, and 70% on CursorBench, a measure of how well models perform inside actual coding workflows. On GPQA Diamond, a test of graduate-level scientific reasoning, it reached 94.2%. VentureBeat reports the new model beats OpenAI's GPT-5.4 and Google's Gemini 3.1 Pro across all four benchmarks Anthropic ran for comparison: agentic coding, scaled tool-use, agentic computer use, and financial analysis.

The practical improvements go beyond scores. Opus 4.7 handles images at higher resolution, up to 2,576 pixels, and Anthropic says instruction-following is noticeably tighter. For developers running long autonomous tasks, the kind that take 30 minutes and involve dozens of files, the model's agentic reliability has improved enough that Anthropic is positioning it as the right choice for work you'd previously have hesitated to leave unattended.

An updated tokenizer is included, and it's worth knowing: it can increase input token counts by 1.0 to 1.35 times compared to earlier versions. If you're running cost projections based on Opus 4.6 usage, they'll need updating. Pricing stays unchanged at $5 per million input tokens and $25 per million output tokens. The model is live on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.

Where it stands in the broader race

The frontier model race has been unusually competitive since early 2026. GLM-5.1 briefly topped the coding leaderboard on SWE-bench Pro. GPT-5.4 led several evaluations earlier this month. Opus 4.7 is the best publicly available model today. That sentence has had a different subject every few weeks this year, and there's no particular reason to think that stops now.