Epoch's first-place ranking for GPT-6 Astra rests on a single coding score

Epoch's own per-benchmark table explains the split. Astra leads on math, knowledge, and puzzles, Fable 5.1 holds the top mark on nearly every coding test, and Epoch has logged exactly one coding score for Astra, from a run at medium reasoning effort ^[17]. Artificial Analysis weights knowledge, coding, and text comprehension ^[9]. A composite is a weighting of somebody else's test mix, and these two do not share one.

The per-token sticker is the wrong unit to decide on. According to the-decoder, OpenAI charges two and a half times as much per unit of processed text as it did for Sol, which works out to roughly 75 percent more per task ^[14]. Those two figures pin the token count: 1.75 divided by 2.5 is 0.70, so a task now consumes about 70 percent of the text it did on Sol ^[23]. The saving comes from step count. Astra needs a third of Sol's compute steps and a fifth of Opus 5's ^[1], and on the Coding Agent Index it scores 67 at roughly a third of Sol's token usage, behind Fable 5.1 at 70 ^[3].

So the 2.5x transfers in full only to work with no steps to save. A single-shot long-context job absorbs the whole increase, and long-context reasoning is one of the places Astra slipped, along with banking support, SciCode, and about 80 Elo on GDPval-AA v2 ^[16]. An agent that loops many times per ticket keeps most of the discount: on coding, Artificial Analysis has Astra level with Claude Fable 5 at less than half the cost per task ^[2].

ARC-AGI-3 shows the same mechanism at its extreme. On the standard scaffold the run cost $49,791 with reasoning off and scored 35.2 percent; at maximum reasoning it cost $26,098 and scored 62.7 percent ^[12]. That is $1,414 per point against $416, a 3.4x improvement in cost per point from turning the dial up ^[22]. ARC Prize attributes it to Astra finishing the games in fewer moves, which means fewer model calls and fewer tokens ^[12]. One setting breaks the pattern. The "low" level scores 17.5 percent, below running with no reasoning at all, and ARC Prize offers no explanation ^[13]. A dial whose second notch is worse than off is one you measure per workload rather than tune by intuition.

The 99.9 percent OpenAI reported came from its own harness, which keeps reasoning chains alive between requests and automatically summarizes long runs ^[25]. ARC Prize measured those runs at about 3.66 times faster with 49 percent fewer tokens, across 167 game-reasoning pairs both setups solved ^[11]. Both features are client-side design choices, so the gap is partly reproducible, but 49 percent is the saving ARC observed on one benchmark's shared subset rather than a number you inherit by bolting on a summarizer.

Epoch applies the same discipline to its Erdos result. Astra was the only model to close two of 68 open problems with Lean-verified proofs on a budget of $300 per attempt, and three further solutions from non-standardized runs costing more than $220,000 were excluded from the score ^[21]. That excluded budget is more than 700 times the standardized per-attempt allowance ^[24]. Chollet's forecast revision rests on the efficiency side of the same benchmark, where Astra beat average human efficiency for the first time and he called the progress twice as fast as he expected ^[19][20]. The measurement that transfers is the one you run on your own mix, at the effort level you would actually ship.