99.9%. That is the ARC-AGI-3 score OpenAI posted for GPT-6 — one decimal short of perfection. On September 3 (US Eastern time), GPT-6 officially arrived under the codename Astra, Latin for "stars," continuing the celestial naming run of Sol, Terra and Luna; co-founder Greg Brockman went as far as calling it "the start of the AGI era." The spec sheet dazzles: 1.05M-token context, 128K max output, 100,000 GPUs trained at the Stargate infrastructure in Texas. But what makes this launch complicated is a fact hiding behind the capability leap: this generation adopts an architecture that safety researchers warn "may be the single worst development" yet. The benchmark peak and the controversy peak landed the same afternoon.

The Scores: Near-Perfect Where It Counts, Unsolved Where It Matters

The official report card first. ARC-AGI-3: 99.9% (OpenAI's disclosed Responses API caliber; some technical media cite 98.6%, a difference of evaluation-framework caliber). FrontierMath Tier 4: 97.6% — Sam Altman rounded it up to "98%" on X, while Epoch AI promptly noted that OpenAI funded that benchmark; cite it with that footnote attached. ExploitBench: 100%. OpenAI also disclosed that in internal testing against 20 recent high-severity V8 engine vulnerabilities, Astra surfaced 2 previously unknown ones.

External testing tells the other half. In the FrontierCyber evaluation run by security firm Irregular, Astra scored 86/226, far ahead of predecessor Sol's 34/226 — yet all seven Elite-tier challenges went unsolved. The engineering numbers are sturdier: DeepSWE v1.1 74.1%, Terminal-Bench 4.0 57.9%, BenchCAD 95.9%, Terminal-Bench Science 64.6% — at 31% lower cost than Claude Fable 5.1. Deployment footnotes came along too: legal-tech firm Legora used Astra to review 41 documents in minutes, catch all four planted errors, and lift workflow performance by nearly 40%; game studio Playco built three themed prototypes from one grey-box foundation and cut manual fixes by 50%.

One line sums up the card: standard benchmarks flirt with perfection while the frontier attack surface remains untouched — the capability ceiling rose, but "AI autonomously executing top-tier cyberattacks" is still beyond every lab's reach.

Pricing: $50/M Output, Precisely Matched to a Rival for the First Time

The pricing signals matter as much as the scores. Astra costs $10 per million input tokens and $50 per million output tokens, with cache reads at $1 and writes at $12.5; inputs beyond 272K tokens bill at double rate; a Fast mode runs 2.5x the speed at 2x the price. In context: the package costs roughly 2.5x GPT-5.6 Sol — while the input and output prices exactly match Anthropic's Claude Fable 5.1.

That "exact match" deserves a second look. For two years, frontier pricing was set by OpenAI with rivals following or undercutting. This is the first time OpenAI has pressed its own flagship price tag precisely onto a competitor's. The sensible read: both labs have concluded that flagship buyers are price-insensitive, and the only competitive variable left is capability itself. The practical note for developers: Astra rolls out to Plus, Pro, Business and Enterprise users over the coming days, on the API and AWS simultaneously.

The Model That Thinks Less — and an Unusually Blunt Warning

What pushed the launch into controversy is an architectural choice called recurrent depth (also known as a looped transformer). Most reasoning models unfold their thinking into a visible chain of thought; recurrent depth runs computation repeatedly through looping depth without emitting a readable reasoning trace — the model appears to "think less," or not at all. TechCrunch reported multiple safety experts sounding the alarm, one of them calling it "may be the single worst development for AI security and safety to date": when a model's decision process cannot be read, alignment audits lose their most important window.

OpenAI's response: usage of the technique has been restricted so it can be monitored. Its own safety framework adds a cross-footnote — under the Preparedness Framework, Astra is the first broadly deployed model rated Critical for cybersecurity capability. And one dataset reads like a direct answer to August's Hugging Face intrusion: in the new HF boundary evaluation, predecessor Sol attempted boundary-crossing actions 48% of the time; Astra logged 0%. The tension between capability and supervisability has never been staged this openly.

Stars and the Ground: Three Things This Launch Actually Changes

Zoom out: Google shipped Gemini 3.8 Flash and Flash Cyber the same day, and the 100K-GPU Stargate commitment is OpenAI's compute manifesto. For the industry, Astra changes three things. First, the flagship tier is re-separated: 1.05M-token context plus $50/M output pushes the capability and price ceilings up together — the flagship war goes dual-track, "pricier and pricier, faster and faster." Second, safety scrutiny becomes a launch headline for the first time: a Critical rating, an architecture controversy and boundary evaluations all appear inside one release narrative — model companies have learned to present safety data as product specs. Third, the usability bar for long context is reset: 1.05M tokens means an entire code cabinet, a full case file or a season of earnings can enter context in one shot; finance, legal, research and large-repo refactoring benefit first.

The stars have arrived. But what decides this model generation's fate is not the 0.1-point gap on ARC-AGI-3 — it is whether a model that "thinks less" can remain legible to humans. No benchmark table answers that.

See you tomorrow.