MHR

Real cost vs intelligence

No-one actually pays subscription pricing, so benchmarks should not measure performance per api pricing, but rather through subscription pricing.

Points are filtered; Pareto lines compare all companies.
Cost axis
Plan cap
USD / month
Applies to subscription fees. Estimates assume full quota use.
Loading chart…
Global Pareto · 6–8 Oct 2026
Explore a model
Logos
AlibabaAnthropicApodex (initials)Arcee AICeleris (initials)DeepSeekGoogleIBMInceptionInclusionAIKimiLongCatMetaMiniMaxMistralMultiverse Computing (initials)NVIDIAOpenAISpaceXAIStepFunTencentThinking Machines (initials)UpstageXiaomiZ AI

Logos by LobeHub.

Method

Since no-one actually pays API pricing for coding tasks, pricing references to API costs are deceiving, since nearly all subscription providers offers steep discounts compared to their API pricing.
We show two paretos, one based on the API pricing (blue), and one based on the subscription pricing that developers are actually expecting to pay. Please note that subscription pricing is assuming that the entirety of the subscription limits are hit. This data is then scaled by how token-efficient each model and reasoning effort is, and measured on the metric of per 1000 Artificial Analysis charts.

Assumptions & snapshot

Intelligence uses AA Index v4.3.2. Estimated plan cost = benchmark API cost × (effective plan token price ÷ calibration API token price). Using half the quota roughly doubles the effective cost. The plan cap chooses the cheapest eligible plan; company filtering preserves global Pareto lines.

Model data checked 8 October 2026.

Data: Real API Pricing & Artificial Analysis · Download JSON