Meta produced the 42% gap without touching its rate card [16]. Token prices held at the Muse Spark 1.2 levels: $1.25 per million input tokens, $4.25 per million output, and $0.15 per million for cached input [8]. Cached input therefore sits at 12% of the standard input price, an 88% discount on precisely the line item that moved, since Artificial Analysis attributed the entire per-task increase over version 1.2 to roughly 57% more input tokens against about 8% more output [3][5]. Gemini 3.8 Flash sat nearest in the same comparison at 59 points for $0.58, about 5% more per task for two fewer points, so the sub-dollar band holds more than one vendor [3][7]. What a buyer controls here is prompt shape and cache hit rate, not the price list.
The two available measurements of token use point in opposite directions. Meta's engineers measured about 20% fewer tool calls and about 25% fewer tokens for 1.3 than for 1.2 [12], while the index measured input use per task rising by more than half [5]. Both can hold on different task mixes, and implicator.ai notes that Meta's own comparison table runs the max variant on Meta's harness, which is not the same class of evidence as the independent index [10]. That leaves one figure measured the same way across all four vendors, the index's cost per task, and it is the only one a procurement conversation can stand on.
A skeptic would say a tie on a nine-evaluation composite is not a tie on any particular workload, and the record supports the objection [9]. Meta's table shows the split: Muse Spark led GPT-5.6 Sol on SWEAtlas CodeBase QnA and trailed it on DeepSearchQA and the Agentic IF Index [13]. What survives the objection is where the gains landed. Tau3-Bench Banking rose from 35% to 47% and Terminal-Bench 2.1 from 80% to 85% between the two releases [11], a twelve-point and a five-point move in multi-step tool use [5], which is the shape of the agent work most buyers are now repricing.
The top of the board sits behind a partner agreement. Max scored 62, one point above xhigh, and remains limited to Meta partners with no disclosed price [7], so the highest score a general customer can transact is 61 at 55 cents [8]. Mark Zuckerberg described the release on X as "frontier performance almost too cheap to meter" [14], and the same release costs 37.5% more per task than the version it replaced [2] while giving back four points on AA-LCR [6]. Both readings are accurate, and they describe the same operating condition: the price floats with model behaviour instead of holding to a posted rate. Repricing agent workloads on this evidence books a real saving this quarter and hands next quarter a budget line whose variance is set by someone else's token accounting.
