Start with the break-even, because it is a single number. A rate card that rises 2.5x is neutral only if the same job burns 40% or less of the tokens it used to [1]. Everything in the Artificial Analysis writeup sorts around that threshold.
The Codex coding-agent runs clear it. A third of the predecessor's tokens at 2.5x the price works out to about 0.83 of the old bill, roughly 17% cheaper per task [2], and Artificial Analysis measures cost per task as about level with GPT-5.6 Sol at max effort, with two more index points [6]. The gap between that arithmetic and their measurement is the part a token ratio hides: input, cached input and output are priced separately, and the mix moves.
The Intelligence Index does not clear it, and the way it misses is instructive. Output tokens fall about 10% at max effort [8]. Price 90% of the old output volume at 2.5x and you would expect about 2.25x the cost, or 125% more; Artificial Analysis reports 75% more [9], which implies the blended token bill actually fell to roughly 70% of the predecessor's [3]. The efficiency gain is real and larger than the output-token figure alone suggests, but it does not reach 40%.
One caution about the coding scores. Astra's 67 was measured in Codex, Fable 5 in Claude Code, and Muse Spark 1.3 in Muse Code [3], with Fable 5.1 in Claude Code leading on 70 [4]. An index row is a model and a scaffold together, so the threefold token reduction belongs to that pair rather than to the weights alone, and a team swapping models inside its own harness does not automatically inherit it.
The result worth the most attention is the hallucination one. On AA-Omniscience the max-effort hallucination rate drops from 92% to 51%, a 45% relative reduction [5], and accuracy rises 4 points at the same time [10], which rules out the usual trade where a model lowers its error rate by declining to answer more often. What it does not tell you is whether the remaining half lands in your domain.
Elsewhere the ledger is mixed. AA-Briefcase, which runs multi-week projects across thousands of source files, gains about 80 Elo on analytical quality while presentation quality falls, with GPT-5.6 Sol still leading that sub-score [11]. GDPval-AA v2, adapted from OpenAI's dataset covering 44 occupations, drops about 80 Elo [13]. Those two 80s sit on different scales and do not cancel. Humanity's Last Exam gains 6 points [12], while customer support, scientific Python and long-context reasoning each slip 2 to 3 points [14].
The read I would commit to, conditioned on these harnesses: the upgrade pays for itself where an agent loop is the token consumer and the score holds flat or slightly better. On the intelligence side, an equal score of 61 against Fable 5.1's 66 [7][4] is now bought at a higher rate, which makes that purchase a decision about hallucination rate rather than about capability.
