Anthropic released Claude Opus 4.7 on April 16, and the improvement in coding capability has surprised many — production code repair tasks are three times better than the previous generation, and SWE-bench Pro jumped from 53.4% to 64.3%.
Three benchmark scores worth watching
Here are the key numbers:
| Benchmark | Opus 4.6 | Opus 4.7 | Change |
|---|---|---|---|
| SWE-bench Pro | 53.4% | 64.3% | +10.9% |
| CursorBench | 58% | 70% | +12% |
| Rakuten-SWE-Bench (production tasks) | Baseline | 3x | Significant improvement |
SWE-bench Pro is now widely regarded as a difficult coding benchmark. At 64.3%, it surpasses GPT-5.4 released in March and Google's Gemini 3.1 Pro from February. In Anthropic's internal tests across 93 coding tasks, Opus 4.7 improved 13% over 4.6 and solved four additional problems that 4.6 could not handle at all.
The Rakuten figure is particularly telling — "production code repair tasks are three times the previous generation" refers to bug fixes in real enterprise codebases, not synthetic datasets. Numbers derived from real production code are more meaningful than lab-generated synthetic benchmarks.
Anthropic also ran its own Finance Agent Evaluation and GDPval-AA economic value knowledge work benchmark, achieving current best scores on both.
Vision capability more than tripled, but rarely mentioned
Beyond coding, Opus 4.7 brings significant upgrades in image processing:
- Supports a maximum long edge of 2,576 pixels (approximately 3.75 megapixels)
- More than three times the previous image resolution ceiling of the Claude series
- Notably improved recognition of chemical molecular structure diagrams and complex engineering schematics
This will have a greater impact on researchers, those analyzing technical documents, or reading engineering drawings. The effect may be less direct for coding, but those working on multimodal engineering should test it.
Another easily overlooked change: the tokenizer has been updated. The same text input now generates 1.0 to 1.35 times the number of tokens. That sounds like bad news (more tokens means higher cost), but it actually indicates the model processes information at a finer granularity, which should lead to more solid understanding of long documents and complex code.
Why Anthropic deliberately disabled cybersecurity capabilities
This is the most counterintuitive decision in the 4.7 release: Opus 4.7 actively adds safety guardrails, automatically detecting and blocking high-risk cybersecurity requests.
Wait — Anthropic just released Claude Mythos Preview last week, focused on cybersecurity, and launched Project Glasswing with Amazon, Apple, and Microsoft. How can both things be true simultaneously?
The answer is product tiering:
- Mythos is a controlled model designed specifically for security professionals, accessible only through identity verification via the Cyber Verification Program
- Opus 4.7 is a general-purpose flagship for the public; Anthropic does not want it to become an attack tool anyone can invoke
The logic holds. Concentrating high-risk capabilities in verifiable professional channels rather than opening them to everyone via API — this is a pragmatic security tiering approach.
New features at a glance
Several auxiliary features launched simultaneously:
- Task budgets (public beta): Lets Claude manage its own token consumption pace during long-running tasks, avoiding sudden budget exhaustion at the final stage
- /ultrareview command: Dedicated code review session that flags bugs and design issues, deeper than ordinary review
- xhigh effort tier: Previously only high was available; now xhigh is added, providing finer adjustment between reasoning depth and response speed
- Auto mode expansion: Available to Max Claude Code users
Pricing remains unchanged: API calls at $5 per million input tokens and $25 per million output tokens, identical to Opus 4.6. Platform support: claude.ai, API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry are all live, with model ID claude-opus-4-7.
How big is the gap with competitors
According to Anthropic, Opus 4.7 surpasses GPT-5.4 (released in March) and Gemini 3.1 Pro (released in February) in agentic coding, large-scale tool calling, agentic computer use, and financial analysis.
Of course, benchmarks are just benchmarks. The real-world experience gap will require more evidence once the community starts using it. But the "3x" figure from Rakuten-SWE-Bench alone is already quite convincing for the production code repair scenario.
From Opus 4.6 to 4.7, this is not a minor iterative update — at least in coding tasks, the magnitude of improvement this time deserves serious attention.
Sources: CocoLoop, Introducing Claude Opus 4.7 (Anthropic Blog); Anthropic's Claude Opus 4.7 makes a big leap in coding, while deliberately scaling back cyber capabilities (The Decoder); Anthropic releases Claude Opus 4.7, narrowly retaking lead for most powerful generally available LLM (VentureBeat)