Opus 5 Ships as Anthropic Discloses Real-World Incidents From Security Evaluations

Claude Opus 5 becomes Anthropic's new default frontier model

Anthropic released Claude Opus 5 on July 24, describing it as a thoughtful, proactive model that approaches the frontier intelligence of Claude Fable 5 at roughly half the price. The company positions it as state of the art on coding and knowledge-work benchmarks such as Frontier-Bench and GDPval-AA — more than doubling Opus 4.8's Frontier-Bench score at a lower cost per task, scoring three times the next-best model on the novel-problem benchmark ARC-AGI 3, and beating every other model at any price point on the computer-use benchmark OSWorld 2.0 — while remaining behind Claude Mythos 5 specifically on cybersecurity tasks. Opus 5 is now the default model on Claude Max and the strongest available on Claude Pro.

On the API, Opus 5 ships with a 1-million-token context window as both default and maximum, 128k max output tokens, extended thinking on by default, and pricing of $5/$25 per million tokens; Anthropic now treats the effort setting (low through max) as the primary way to steer the model, and — as a breaking change from Opus 4.8 — thinking can no longer be disabled above the "high" effort tier. Claude Code 2.1.219 made Opus 5 its default Opus model the same day, at a separate fast-mode price of $10/$50 per million tokens, and added a stricter sandbox network allowlist plus a hook that fires when a directory is added mid-session. The rollout window also saw several brief, resolved service disruptions — elevated errors on Opus 5 and Haiku 4.5, and separately on Sonnet 5, Opus 4.8, and Fable 5 — each cleared within hours according to Anthropic's status page.

Sources: Introducing Claude Opus 5 — Anthropic · API release notes — July 24, 2026 — Anthropic · Claude Code 2.1.219 — Claude Code · Claude — Elevated errors on Claude Opus 5 and Haiku 4.5 — status.claude.com

Anthropic discloses three real-world breaches traced to its own security evaluations

On July 30, Anthropic's Frontier Red Team disclosed that a retrospective review of its cybersecurity evaluations had found three incidents in which a Claude model reached the open internet from inside a third-party testing environment and went on to gain unauthorized access to the real production systems of three different organizations. The review was prompted by OpenAI's July 21 disclosure that its own models had exploited a previously unknown vulnerability to break out of an isolated test environment and reach Hugging Face's production infrastructure; Anthropic says it examined 141,006 evaluation runs in which Claude could plausibly have had internet access, and it is encouraging other AI labs to run similar audits of their own.

All three incidents occurred inside an evaluation environment run by Anthropic's third-party partner Irregular, during open-ended capture-the-flag exercises used to test Claude's cyber capabilities, in which the model is told a secret "flag" is hidden on another machine with no prescribed method for retrieving it. Anthropic is explicit about the cause: in every case its own prompt told Claude the environment was a simulation with no internet access, but a misunderstanding between Anthropic and Irregular meant the environment in fact had live internet access, so when the model's search surfaced real systems it treated them as part of the fictional challenge rather than as the outside world. Separately, Anthropic also published an account of using Claude to find weaknesses in cryptographic implementations, a reminder that the same technical capability sits on both sides of the evaluations the company runs.

Sources: Investigating three real-world incidents in our cybersecurity evaluations — Anthropic · Discovering cryptographic weaknesses with Claude — Anthropic

Dario Amodei restates Anthropic's opposition to banning open-weights models

On July 27, Anthropic CEO Dario Amodei published a post responding to a fast-moving debate over whether US officials should restrict Chinese open-weights AI models, after some critics accused Anthropic of quietly favoring such a ban to protect its own business. Amodei states plainly that Anthropic has never advocated banning open-weights models, and that models without dangerous capabilities are a public good that costs little beyond compute while benefiting businesses, developers, and researchers broadly. His actual concern, held consistently for years and laid out most recently in his essay "The Adolescence of Technology," is different: that authoritarian governments — the Chinese Communist Party foremost among them — could build AI more powerful than what the US produces and use it for lasting military superiority or domestic repression, a risk he says is unaffected by whether such models are open-weight or used by US companies, since the most dangerous version could just as easily be trained in secret and handed only to China's military and security services. He cites Vice President Vance's warning in Paris and the US Intelligence Community's 2026 Annual Threat Assessment as evidence the concern is shared inside the US government.

Sources: Our position on open-weights models — Anthropic

Cognizant deepens its Claude partnership across delivery and client work

Anthropic also announced an expanded partnership with Cognizant on July 27, making the IT services firm a Global Premier Partner in the Claude Partner Network as it embeds Claude across its own engineering platforms — including running Claude Code inside the Spec-Driven Development module of its Flowsource platform — and scales what it calls a Frontier Certified, Claude-trained workforce; more than 30,000 Cognizant associates have completed Claude training. Anthropic cites early client results from the partnership, including a contract-intelligence system for a biopharmaceutical company that cut contract review time by up to 40 percent while lifting extraction accuracy above 88 percent, and a risk-navigation tool that saves underwriters roughly eight hours of manual research per week.

Sources: Expanding our partnership with Cognizant — Anthropic