Claude Opus 5.5 charges 5.3 times GPT-6 Sol's price per task for a nine-point benchmark lead

On Zapier's AutomationBench, which runs agents through end-to-end workflows across 47 business tools, Claude Opus 5.5 led the September 28 table at 42.47% for $1.44 a task ^[8]. GPT-6 Sol, at its XHigh setting, scored 33.2% for $0.27 ^[8]. The extra $1.17 buys 9.27 points ^[7][2]. Counting failures changes the ratio. If the score is taken as the share of workflows an agent finishes, a finished workflow costs about $3.39 on Opus and $0.81 on Sol, and the premium falls from 5.3 times to about 4.2 times ^[3][1].

Cost per task also moves with the measurement. Artificial Analysis estimates Sol at maximum effort at $1.06 a task on its Intelligence Index, about 3.9 times its Zapier figure in the same month ^[9][5]. The effort settings differ, and so do the tasks. A valuation built on cost per completed task has to name the workload before the figure means anything, because the same model's number moves by a factor of about four between two tables ^[5].

The shift the FT describes could still miss vendors' revenue in several ways. The first is volume. Axios cites the Jevons paradox, that cheaper resources get used more, and Gartner's forecast has agentic inference bills rising even as unit prices fall ^[12][11]. A vendor can lose on price per task and still grow revenue. The second is decay: at Epoch AI's rate, the cost of a given level of performance roughly halves each quarter, so a two-times cost advantage is gone in about one quarter if a rival follows that curve ^[6]. The third is that buyers keep paying for the top model anyway. Ramp's count of buyers favours Anthropic, whose Opus 5.5 is the pricier of the two models on Zapier's table ^[3][8]. The index as reported counts businesses, not spending by model, so that lead could rest on Opus or on Anthropic's cheaper models ^[3].

I think the FT's pattern is right about buyers. The OECD's interviews with 25 organisations found cost savings a common reason for using agentic systems ^[13]. A company routing by cost stops paying $1.44 for work Sol can finish at $0.27, and pays the frontier price only where the extra completion rate changes the outcome ^[1][10]. For an investor, the useful figure is a vendor's cost per completed task on the workloads it sells into, set against how much of the larger bill Gartner expects actually lands with that vendor ^[11]. The view is wrong if Ramp's later data show spending concentrating on frontier models while token prices keep falling, since that would mean buyers take the cheaper tokens and still choose the top model ^[2][3].