Anthropic drops Claude Opus 4.7 with supercharged coding and cyber guardrails

Anyone who has tasked an AI coding assistant with a massive refactor knows the familiar dread: twenty minutes in, it hits a tool error, starts looping endlessly, or hallucinates scaffolding code to cover its tracks.

Anthropic officially rolled out Claude Opus 4.7 across Claude apps, its API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Pricing stays frozen at $5 per million input tokens and $25 per million output tokens, but the model under the hood is dialed squarely into multi-step autonomy and self-verification.

Reliable multi-hour runs over raw benchmark flexes

Opus 4.7 tackles complex software engineering and long-horizon tasks by checking its own logic before surfacing answers. It introduces new knobs like an xhigh effort level alongside task budgets, while bumping multimodal resolution limits up to 2,576 pixels.

Early partner benchmarks show real jumps across production harnesses. Cursor reported its CursorBench score spiked from 58% on Opus 4.6 to over 70% on Opus 4.7. Rakuten logged a 3x increase in resolved production tasks on Rakuten-SWE-Bench. Cognition, the team behind Devin, noted the model can run coherently for hours without stalling, while Notion found tool call errors dropped by two-thirds on multi-step workflows.

Hex pointed out another subtle but critical behavior: low-effort Opus 4.7 roughly matches medium-effort 4.6, and the model now refuses to hallucinate plausible workarounds when underlying data is missing.

Putting the brakes on cyber offensive capabilities

What makes this launch unusual is what Anthropic deliberately took out. Following its Project Glasswing announcement, where the company pledged to test new cyber guardrails on smaller models before rolling out its powerful Mythos-class previews, Opus 4.7 had its high-risk cyber capabilities intentionally reduced during training.

The model ships with automated systems designed to flag and block prohibited high-risk cybersecurity requests. Legitimate security professionals looking to run penetration testing or vulnerability research will have to sign up through Anthropic's new Cyber Verification Program.

Why it matters

We think Anthropic is reading the enterprise room correctly here. Developers do not just need a model that writes snippets; they need an agent that does not spin out of control when left unattended in CI/CD pipelines. Keeping the price identical while drastically cutting down on tool failures and loops makes this an automatic upgrade for technical teams.

It also sets an intriguing precedent for model governance. Instead of shipping maximum horsepower across every vector, Anthropic is actively dialing back specialized offensive capabilities while keeping the defensive reasoning sharp. The catch will be false positives: automated safety filters are notorious for breaking benign debugging, so how smoothly the verification program runs will decide whether security pros stay on board.