The distance between two numbers is the thing a procurement table exists to supply, and that distance moved while neither model did. In the 2:23 p.m. Internet Archive snapshot of the launch post, Astra led Anthropic's Fable 5.1 on FrontierMath Tier 4 (v2) by 9.8 points, 97.6 against 87.8 [9][10][1]. In the 5:17 p.m. snapshot it led by 19.6, 97.6 against 78, exactly twice the original margin, with Astra's own figure untouched throughout [9][10][2]. That shift happened in two hours and fifty-four minutes [8]. The margin now reads 14.6 [3].
An OpenAI spokesperson told Fortune that most evaluations carry noise within a few percentage points depending on the exact checkpoint, scaffold and evaluation run, and that the launch-blog fixes were made so users could make meaningful comparisons [5]. A competitor's score falling 9.8 points and then recovering 5 of them sits well outside a band described as a few [7][9].
The same afternoon halved Astra's reported hallucination rate from 4.2% to 2% alongside four other changed metrics, and cut GPT-5.6 Sol's from 12.2% to 9.4%, before both returned to where they started [6][7]. Read as a ratio, which is how anyone comparing generations reads it, Astra went from 2.9 times cleaner than its predecessor to 4.7 times cleaner and back [4]. On ARC-AGI-3 the embargoed draft sent to media said 98.6% and the live page says 99.99% [12], a 1.4-point move that is comfortably inside OpenAI's stated noise band and simultaneously a 140-fold reduction in residual error [5]. Near the ceiling, a point is not worth what a point is worth in the middle.
The pattern is not uniformly self-serving. Sol's internal ExploitBench score was raised from 5.5% to 11.5%, a doubling that flatters the predecessor rather than the new model [8][6], and Sol's hallucination figure improved in the same pass [7]. More to the point, the flattering numbers did not stick: the hallucination rates are back at 4.2% and 12.2%, and Fable is back at 83% [6][7][10]. A lab willing to walk a number back this way is doing something closer to error correction than a lab that lets an inflated figure ride.
Which leaves the ExploitBench disclosure as the load-bearing item, or rather the more useful version of it: OpenAI says it is investigating reverting Sol's 11.5% because that result reflects a reasoning level not commercially available for Sol [8]. A published figure describing a configuration nobody can buy is an accurate measurement of something else entirely, not a noisy measurement of the product, and no error bar reconciles the two.
What would falsify the sceptical read is disclosure rather than restraint. If each cell arrives with a checkpoint, a scaffold, a reasoning level and a run date, and independent reruns land inside the few-point band OpenAI describes [5], the table becomes usable with a tolerance attached. Absent that, the observed movement of 9.8 points on a comparator [7] exceeds most of the margins a buyer would act on, which makes the vendor table a shortlist filter and leaves the scoring to a harness the buyer pays for.
