OpenAI launches GPT-6 Astra, scoring 99.9% on ARC-AGI-3 and saturating the benchmark

OpenAI's GPT-6 Astra Saturates ARC-AGI-3 at 99.9%, Redefines Frontier

OpenAI launched GPT-6 Astra with benchmark results that effectively close out one of AI's most challenging evaluation frameworks. The model scored 99.9% on ARC-AGI-3, the pattern-recognition benchmark that has defined capability discussions since its introduction. On Frontier Math Tier 4 v2, Astra posted 97.6%. It saturated X-Bladebench entirely at 100%.

The release lands amid a week of unusually coordinated model announcements from major labs. Anthropic released Claude Fable 5.1 (scoring 66 on the Artificial Intelligence Index). Google shipped Gemini 3.8 Flash. Meta released NewSpark 1.3. The timing suggests what one host characterized plainly: lab leaders decided people like AI and launched simultaneously to keep momentum.

But the ARC-AGI-3 result carries real weight. The benchmark, created by Francois Chollet at ARC Prize, was designed as a hard floor—one that required genuine inductive reasoning rather than pattern matching on training data. At 99.9%, the model has essentially solved the task as specified.

The goalpost moves immediately. Chollet responded on the morning of Astra's launch by announcing ARC-AGI-4, which will require models to demonstrate open-ended invention and novel physics understanding. His framing was direct: "We lack evidence to call this AGI yet." The new benchmark will measure whether models can actually discover and formalize new scientific principles.

The result illustrates a recurring tension in frontier AI evaluation. Benchmark saturation happens fast enough now that the moment a test launches widely, labs can direct training toward it. Within months, the ceiling moves. This has sparked real skepticism about whether standard benchmarks retain much signal. One host noted that benchmark trust is "at an all time low and maybe is headed lower."

Yet the practical capabilities implied by Astra's performance—particularly in reasoning and computer use—remain significant. Early benchmark details show strong scores on Reasoning Effort Max (64.6%) and the reasoning effort honeypot at 0% (meaning no exploitative shortcuts detected). OpenAI is positioning it as "the world's best computer use model," a claim that will see rapid real-world testing.

The larger story is not the benchmark number itself but the velocity. Astra launches. Within hours, the benchmark is obsolete by design. The next bar is already drawn. Labs are training toward it. And AI capability gains have become a calendar-based narrative rather than a scientific one.