LLM landscape

Capability per inference cost

How much capability each dollar of inference buys among models currently scored by Epoch ECI. New releases enter after Epoch publishes a score; its publication lag is not fixed. Price is the published API list price per million tokens.

Data as of
Aug 25, 2026
Priced models

54

Providers

7

Price basis

Public list

Capability coverage follows the current Epoch ECI dataset.

No fixed publication lag. New releases enter this chart after Epoch publishes an ECI score. The refresh timestamp records a successful pipeline run; it does not certify same-day coverage of every model release.

Intelligence per dollar

Capability vs inference price

Each dot is a model · log price axis · higher and left-er is better value

Among models currently scored by Epoch ECI, GPT-5.6 has the highest capability score (161.1) at $11/Mtok — about 86× the price of the cheapest option, Gemini 1.5 Flash (130.2). The dashed line marks the cost-efficiency frontier: the most capability available at each price.

← cheaper · more capable ↑ = better value110120130140150160$0.1$0.30000000000000004$1$3$10$30$100Blended price, USD / million tokens (log scale)Capability (ECI score)GPT-5.6GPT-5.6 TerraGPT-5.5GPT-5.6 LunaGPT-5.4 NanoGemini 1.5 Flash
Organisation
MarkerOpen weightsProprietaryTier-estimated priceCost-efficiency frontier

Capability is the Epoch Capabilities Index (ECI), rebuilt by the scheduled dataset refresh. Inference price is the published first-party API list price in USD per million tokens (open-weight models use a representative hosted price); each model carries a source link and an as-of date. The blended price weights input and output 3:1, the common convention for a single headline rate. Some newer point releases that a lab has not separately repriced are shown at their tier rate (Opus $15/$75, Sonnet $3/$15, GPT-5 $1.25/$10, etc.) and flagged as a tier estimate. Models without any public or tier price — and premium 'Pro' variants whose pricing is not a simple tier — are omitted, so the newest releases appear only once their pricing is added to the curated table. 0 priced entries did not match a current ECI-scored model.