OpenAI's GPT-6 Astra system card discloses it is the first broadly deployed OpenAI model to reach the Critical tier for cybersecurity capability, and third-party compilations show large Astra leads on offensive-security benchmarks (SRE-Bench 88.0% vs 12.5%, ExploitBench 100.0% vs 70.0%) while Claude Fable 5.1 retains the SWE-bench Pro coding lead at 81.2%. Benchmark leadership is now split by task category rather than concentrated in one lab, and reports suggest Anthropic may accelerate a release in response. For financial institutions, this fragments the vendor-evaluation picture: cybersecurity capability and coding capability become separate procurement decisions, and the Critical classification adds access-control and sign-off friction in regulated deployments.
213 other tracked developments touching the same claims.