Claude Opus 4.7 Is Here: Smarter Coding, Sharper Vision, and a Few Guardrails

Anthropic just released Claude Opus 4.7, and it’s now generally available across all their products, the API, and cloud platforms like Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry. Pricing stays the same as Opus 4.6: $5 per million input tokens and $25 per million output tokens. Developers can hit claude-opus-4-7 via the API.

Let’s get the headline out of the way: Opus 4.7 is a meaningful step up from Opus 4.6, especially in advanced software engineering. Early testers are reporting they can hand off their hardest coding tasks—the kind that used to require constant babysitting—and trust it to deliver. The model apparently catches its own logical mistakes during planning, verifies its outputs before reporting back, and handles long-running tasks with more consistency than before.

Vision also got a real upgrade. Opus 4.7 can see images at higher resolution, which means better reading of complex diagrams, chemical structures, or technical docs. Early adopters like Solve Intelligence are already using it for life sciences patent workflows—drafting, prosecution, infringement detection. That’s a niche but telling use case.

One thing that stands out from the early tester feedback is how Opus 4.7 handles uncertainty. Hex, a data platform, noted that the model correctly reports when data is missing instead of inventing plausible-sounding but wrong answers. That’s a big deal for anyone who’s been burned by LLMs confidently hallucinating. It also resists what they call “dissonant-data traps” that even Opus 4.6 fell for.

On Hex’s internal 93-task coding benchmark, Opus 4.7 improved resolution by 13% over Opus 4.6, including four tasks that neither Opus 4.6 nor Sonnet 4.6 could solve. That’s not just incremental—that’s crossing thresholds.

Another tester, Cognition (the folks behind Devin), said Opus 4.7 “takes long-horizon autonomy to a new level” in Devin. It works coherently for hours, pushes through hard problems instead of giving up, and unlocks investigation work they couldn’t reliably run before. If you’ve ever watched an agent stall out on a multi-hour task, you know how valuable that is.

But here’s the interesting wrinkle: Opus 4.7 is not Anthropic’s most powerful model. That title still belongs to Claude Mythos Preview, which they announced last week alongside Project Glasswing. Mythos is broader and more capable, but they’re keeping it on a short leash due to cybersecurity risks. Opus 4.7 is effectively the testbed for the safeguards they want to eventually deploy on Mythos-class models.

Specifically, Opus 4.7 has reduced cyber capabilities compared to Mythos. Anthropic says they experimented with differentially reducing these capabilities during training. The model ships with safeguards that automatically detect and block requests indicating prohibited or high-risk cybersecurity uses. If you’re a legitimate security professional—vulnerability research, penetration testing, red-teaming—you can apply for their new Cyber Verification Program to get access.

This is a smart move. Rather than holding back the entire model or releasing it without guardrails, they’re shipping a slightly nerfed version with real-world safety testing built in. It’s pragmatic, even if it means power users might grumble about the limitations.

Testers across the board are impressed. Replit called it an “easy upgrade decision.” A financial technology platform noted that for serving millions of consumers and businesses, the combination of speed and precision could be “game-changing.” Another tester said Opus 4.7 is “the strongest model Hex has evaluated.”

On benchmarks, Opus 4.7 tied for the top overall score across six modules at 0.715 on one internal research-agent benchmark, with the most consistent long-context performance of any model tested. On General Finance, it scored 0.813 versus Opus 4.6’s 0.767. Deductive logic, where Opus 4.6 struggled, is now solid.

Is it perfect? No. It’s less broadly capable than Mythos, and the cyber limitations might frustrate some developers. But for most practical work—coding, agent workflows, vision tasks, multi-step reasoning—this looks like the best Claude model you can actually use today. And the price hasn’t changed, which is always welcome.

If you’re already in the Claude ecosystem, go test it. If you’re not, this might be the moment to give it a serious look. The gap between “can do simple tasks” and “can handle my hardest work reliably” is finally closing.