Anthropic releases Claude Opus 4.6

Anthropic released Claude Opus 4.6, describing it as its strongest model to date on coding, agentic and long-context tasks. The release included a 1-million-token context window in beta on the Claude API, which Anthropic said sharply improved long-context retrieval — 76% on an eight-needle test at 1 million tokens, against 18.5% for Sonnet 4.5 — alongside gains on agentic coding benchmarks such as Terminal-Bench 2.0 and on Humanity’s Last Exam. Pricing stayed at $5/$25 per million input/output tokens, with a premium rate above 200,000 tokens of context.

Two accompanying reports drew more attention than the capability claims. A 53-page sabotage risk report assessed whether Opus 4.6, given real access to Anthropic’s own workplace systems, could quietly interfere with code, decisions or infrastructure in ways that raised the chance of later harm. It mapped eight specific pathways through which a misaligned model could cause damage and concluded the risk was “very low but not negligible” — the model showed no steady hidden drive toward sabotage across many tests, and in a deception-style evaluation succeeded at a hidden side task only around 18% of the time, but Anthropic did not rule the risk out entirely.

Separately, Anthropic reported that Claude Opus 4.6, working without specialised instructions or custom scaffolding, had found and validated more than 500 high-severity, previously unknown vulnerabilities in open-source software, with patches already landing in some affected projects. The company said every bug was extensively validated before disclosure, and that the model’s approach — reading code and reasoning about it the way a human security researcher would, rather than fuzzing inputs at random — differed in kind from prior automated vulnerability-discovery tools. Anthropic introduced new cybersecurity-specific detection probes alongside the release, framing the same capability as both a defensive tool and a misuse risk requiring monitoring.