Anthropic released Claude Opus 4.7 on Wednesday, and on paper it looks like a straightforward upgrade. Better at coding. Better at following instructions. Better at long-horizon agentic work. But the most interesting thing about this model is not what it gained. It is what Anthropic deliberately took away.

The headline numbers are real. On SWE-bench Pro, the industry's gold standard for autonomous coding, Opus 4.7 scores 64.3%. That is up from 53.4% on Opus 4.6 and ahead of OpenAI's GPT-5.4 at 57.7%. On the GDPVal-AA knowledge work evaluation, it posted an Elo score of 1753, comfortably ahead of GPT-5.4 (1674) and Gemini 3.1 Pro (1314). For developers building agentic workflows, the model delivers 14% better performance on complex multi-step tasks while using fewer tokens and producing a third of the tool errors.

The Cyber Nerf Is Deliberate

Here is where it gets interesting. Opus 4.7 actually scored lower than its predecessor on cybersecurity vulnerability reproduction: 73.1% versus 73.8% for Opus 4.6. That is not a bug. Anthropic says it "experimentally tried to reduce certain cyber capabilities differentially during training." New safeguards automatically detect and block requests that indicate prohibited or high-risk cybersecurity use.

This is the direct fallout from Project Glasswing, the cybersecurity initiative Anthropic launched earlier this month around Claude Mythos Preview, the model they have explicitly called too dangerous for general release. Mythos remains restricted to a small group of enterprise partners doing defensive security work. Opus 4.7 is the first test case for whether Anthropic can surgically reduce dangerous capabilities while keeping everything else sharp.

Three Times the Vision, Same Price Tag (Sort Of)

The technical upgrades beyond coding are significant. Image resolution has tripled: Opus 4.7 processes images up to 2,576 pixels on the long edge (roughly 3.75 megapixels), compared to the previous ceiling. On document reasoning benchmarks, accuracy jumped from 57.1% to 80.6%. Visual acuity tests went from 54.5% to 98.5%. For anyone building computer-use agents that need to navigate dense interfaces, this is a major unlock.

Pricing stays at $5 per million input tokens and $25 per million output tokens. But there is a catch: a new tokenizer maps the same text to up to 1.35x as many tokens, and the model generates more output at higher effort levels. Anthropic is marketing unchanged prices while the real cost per request can climb meaningfully. A new "xhigh" effort setting slots between "high" and "max," and task budgets in public beta let developers cap token spend for autonomous agents.

The Mythos Shadow

Every piece of Opus 4.7 marketing circles back to Mythos. Every benchmark chart includes Mythos Preview numbers sitting above the rest. On SWE-bench Pro, Mythos scores 77.8% to Opus 4.7's 64.3%. On nearly every test, Mythos is the clear ceiling.

As Gizmodo put it, the release of Opus 4.7 reads like "a promotion for Claude Mythos Preview." That framing is hard to argue with. Anthropic is simultaneously telling you this is their best generally available model and making sure you know something much better exists that you cannot have.

The strategic play is clear. Anthropic is using Opus 4.7 as a proving ground for differential capability control. If they can successfully ship a model that is excellent at coding and reasoning but deliberately limited in cybersecurity exploitation, it builds the case for eventually releasing Mythos-class models to the public. The company says it has no plans to make Mythos Preview generally available, but everything about this release is designed to learn how to eventually do exactly that.

Opus 4.7 is available now across all Claude products, the API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. It is also rolling out on GitHub Copilot.

First reported by Anthropic. Additional reporting by CNBC, VentureBeat, The Decoder, and Gizmodo.