The ratio between the two ARC-AGI-3 scores is 1.59 [1], and the mechanism behind that spread is not mysterious: a developer-run evaluation environment comes with tailored API access and custom prompting, while ARC Prize's Standard harness is infrastructure OpenAI did not configure [20]. The more interesting version of the objection is that a production deployment looks more like OpenAI's harness than ARC Prize's, because nobody licenses a frontier model in order to run it bare, which would make 62.7% a floor rather than a correction [6]. The floor is the part a buyer can verify.
On Artificial Analysis's Intelligence Index v4.1.1, Astra came in at 61.2 against GPT-5.6 Sol's 60.9 [13], a gain of 0.3 points, call it half a percent [2], and it sits 4.5 points behind Claude Fable 5.1's 65.7 [3]. On Humanity's Last Exam it scored 57.2% with tools against Sol's 65.0% [14], which is 7.8 points down, 12% in relative terms [4]. The gains that survive independent scoring are narrow and they land where OpenAI spent: Terminal-Bench 4.0 at 57.9% against Sol's 37.3%, a factor of 1.55 [15][5], FrontierMath Tier 4 at 97.6% and ExploitBench at 100% [16], and 72.6% on OSWorld 2.0 in roughly 47% less time per task [17].
That time saving is internally consistent with the 1.9x figure quoted for Codex, since one divided by 0.53 is 1.89 [7]. Then the safety system takes its cut. At a 20% compute overhead [9] you are paying something like 1.2 units of compute for one unit of work, so a 1.9x speed advantage nets out near 1.58x [6], before pricing the interruptions to legitimate work that OpenAI has already conceded the monitor will cause [9]. The launch material carries no pricing, so whether that overhead lands on the invoice or inside OpenAI's own margin is simply not disclosed.
Which brings the AGI declaration [4] up against the company's own paperwork. The charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work [10], OpenAI maintains GDPval to measure precisely that, and GDPval does not appear in the Astra launch materials [11]. Artificial Analysis ran its own variant and reported roughly 80 points gained on long-horizon knowledge work alongside declines in categories including banking support and scientific coding [12]. Brockman, for his part, told reporters there was no obvious recognisable moment of the kind the company once expected [19], and ARC Prize called the result a step-function change and a milestone worth celebrating while declining to call it AGI [7].
Astra also ships switched off, with enterprise workspace administrators required to enable it manually [3], which puts the decision in the hands of the person who will field the complaints when the monitor interrupts a job, and who is doing that instead of shipping something else that quarter.
My read is that what is for sale is throughput in terminal and computer-use work, announced as general intelligence; the honest counter is that harness engineering is real engineering, and a well-built pipeline may genuinely sit closer to 99.9 than to 62.7 [5][6]. The thing that would break my read is GDPval numbers for Astra showing gains across categories rather than the split profile Artificial Analysis found [11][12]. Until those exist, the auditable claims are 62.7%, 0.3 points, and a 20% tax.
