OpenAI's latest model blows past GPT-5 on Epoch AI's composite index and becomes the first to reach 'critical' cybersecurity tier
OpenAI’s Astra just posted a 169 on the Epoch Capabilities Index, the composite benchmark maintained by independent research organization Epoch AI. That score aggregates performance across more than 37 individual evaluations spanning math, coding, science, and agentic tasks, and it represents the highest number any AI model has achieved on the index.
To put that in perspective: Claude 3.5 Sonnet sits at 130 on the same scale, and GPT-5 scored 150.
What the numbers actually look like
The raw benchmark results behind that composite score are striking on their own. Astra recorded a perfect 100% on ExploitBench, a cybersecurity evaluation that tests a model’s ability to identify and exploit software vulnerabilities. It hit 99.9% on ARC-AGI-3 when run through a Provider Adapter harness, a test designed to measure abstract reasoning and pattern recognition. And it scored 98% on FrontierMath Tier 4, the most demanding tier of a mathematics benchmark built to challenge frontier-level models.
Astra also posted strong results in math, puzzle-solving, and coding suites, though the specific scores on those individual evaluations weren’t broken out in the same detail.
The cybersecurity milestone that matters most
Perhaps more consequential than the headline ECI number is a quieter achievement buried in the results. Astra is the first OpenAI model to reach the “Critical” tier under the company’s own Preparedness Framework for cybersecurity. That designation means Astra has demonstrated the autonomous ability to discover zero-day vulnerabilities and execute exploit chaining without human intervention.
OpenAI’s “Path to Astra” update, published on September 1, 2026, detailed these capabilities alongside new safeguards. Among them: enhanced chain-of-thought monitoring, a technique that allows safety teams to inspect the model’s reasoning process in real time. The company indicated a staggered rollout with particular emphasis on responsible usage in sensitive areas, suggesting that full cybersecurity capabilities won’t be available to all users on day one.
What composite benchmarks tell us, and what they don’t
The ECI has become increasingly important as a measurement tool precisely because the AI industry has a benchmark saturation problem. When a model scores 99% on a test, and its successor scores 99.5%, the improvement is real but nearly invisible at the individual level. By combining 37-plus evaluations into a single rescaled number, Epoch AI gives researchers and observers a cleaner signal of overall capability growth.
The choice of reference points matters here. Pegging Claude 3.5 Sonnet at 130 and GPT-5 at 150 creates a familiar-feeling scale, though it’s worth noting this is an arbitrary normalization, not an IQ test. The numbers enable comparison, not absolute measurement.
