Anthropic ships a price dial with its new model, and that is now the buying decision

The cost advantage does not travel between workloads. On CursorBench 3.2 at maximum effort, Anthropic reports Opus 5 landing within 0.5% of Fable 5's peak score at half the cost per task ^[7]. On OSWorld 2.0, a computer use benchmark, it says Opus 5 beats Fable 5's best result at just over a third of the cost ^[9]. Those two are a 2x and a roughly 3x saving on the same comparison ^[16]. A savings model built on the headline "half the price" ^[1] will therefore be wrong in both directions, and which direction depends on what your teams actually do all day.

The price side is flat. Anthropic says Opus 5 costs the same as Opus 4.8 ^[5] while more than doubling its predecessor's Frontier-Bench v0.1 score at a lower cost per task ^[6], which puts performance per dollar on that benchmark at better than 2x for an unchanged list price ^[17].

The tiers also do not stack. The near-parity result comes from max effort on one benchmark ^[7], while the claim that even the cheapest setting beats every rival comes from Zapier AutomationBench, where Anthropic also reports a pass rate around 1.5x the next-best model at the same cost per task ^[10]. Those are different dials on different tests ^[19]. A team that reads the parity number and then runs its whole estate at low effort has bought neither result.

The verification problem sits underneath all of it. Every figure here, including the FrontierCode 1.1 comparison reported through an early-access customer ^[15], comes from Anthropic's own announcement ^[18]. The vendor is also the one telling you where it loses: it concedes cybersecurity tasks to Mythos 5 ^[3] while claiming state of the art on coding and knowledge work evaluations ^[2]. That is a useful admission, because it means "best model" is scoped to a task class, and making Opus 5 the standing default on Claude Max ^[12] carries an exception the supplier named before any buyer did.

The qualitative material is more interesting than the charts for tier selection. Anthropic describes a Frontier-Bench task where the model, denied any way to view a drawing, wrote its own computer vision pipeline to extract the geometry, and says no competing model in the same setup solved it in five attempts ^[14]. Persistence of that kind is exactly what a low effort setting is meant to ration. The research gains follow the same pattern of narrow, checkable deltas: 10.2 percentage points over Opus 4.8 on inferring molecular structures from spectroscopy data, 7.7 on predicting the functional effect of protein sequence variation ^[13].

So the number that matters is not on the vendor's chart. It is your own cost per completed task at each effort setting, on your own work, and nobody outside your organisation can compute it.