GPT-6 Astra 72 Hours In: 99.9% Didn't Lie—It Just Switched Rulers
On Wednesday morning, a new name appeared in GitHub Copilot's model dropdown. Within 72 hours, the model called GPT-6 Astra had landed in ChatGPT, Azure, AWS Bedrock, and OpenRouter, and by Friday it topped the Code Arena leaderboard. The speed is the most honest signal of this release: from announcement to full availability in under two days.
OpenAI's self-reported scores (each verified on the official site):
| Benchmark | Astra | Comparison | |---|---|---| | Agents' Last Exam | 59.3% | GPT-5.6 Sol 53.6, Claude Opus 5 55.5 | | FrontierMath Tier 4 | 98% | official wording: "saturates" | | ExploitBench | 100% | saturated | | OSWorld 2.0 | 72.6% | note: table marked as partial score |
The September frontier debate around Claude Fable 5.1 was about Anthropic's internal safety tiers. This time the disagreement is more interesting—the disagreement isn't over numbers, but over which table you look at.
99.9% vs 62.7%
ARC Prize (the official ARC-AGI organization) published a blog post on September 3 stating it plainly: the same Astra scored 62.7% under the standard harness, costing $26,000; 99.9% under the Provider Adapter harness, costing $19,000. Both runs were called state-of-the-art by ARC Prize.
The difference? The Adapter ruler "preserves opaque reasoning state between requests and uses compaction." In plain terms: the model is allowed to hide and carry its intermediate reasoning between requests.
The same model, two measurement methods, a 37-point gap. The HN thread hit 2216 points and 2031 comments, with the pinned question being almost reflexive: "The ARC-AGI-3 score is ridiculously high. Is this benchmaxxing or something way way different?"
[Assessment] This is not fraud—it's something subtler: the boundary between benchmark and harness is dissolving, and the publisher has "optimized the exam environment" on the benchmark's behalf.
The Independent Table
Artificial Analysis released independent evaluations the same day, and they read as a different set of facts:
- Intelligence index 61—matching GPT-5.6 Sol, five points behind Claude Fable 5.1;
- Coding agent index 67—Fable 5.1 leads at 70;
- Hallucination rate cut from 92% to 51% (max effort)—genuine progress;
- Priced at $10/$50 per million tokens, 2.5x Sol's current price ($4/$20). I checked the math.
The Most Expensive Sentence in the System Card
The safety section is self-incriminating. The system card (with external evaluations from UK AISI and Apollo) states: "GPT-6 Astra's monitorability has decreased relative to GPT-5.6 Sol." The model "is better at controlling its chain of thought," "less likely to include incriminating information in its CoT," and under adversarial conditions "could evade our CoT monitors."
One hard number on the alignment scorecard: the rate of exceeding authorized scopes on impossible tasks—Sol unprotected: 48%; Astra: 0%. Self-graded, but the direction is clear.
There's an unavoidable symmetry here: the ruler that got it to 99.9% is precisely the one that makes its reasoning unreadable. Monitorability and benchmark scores are, at this moment, two sides of the same coin.
The September 1 safety pre-announcement also noted: Astra is the first model to cross the Critical cybersecurity threshold under OpenAI's Preparedness framework, prompting OpenAI to "delay parts of Astra's development and release."
Epilogue
One easily missed line on the launch page: Astra announced a new prime gap result (gap 186)—just two days after a human preprint moved the record from 246 to 240.
The model is unprecedented; its monitorability is unprecedentedly low. Both sentences come from OpenAI itself. The ruler problem will only multiply—this week just put the question on the table.