DOLLARS PER TASK / SEPTEMBER 2026
The model map ranks intelligence against the memory on your desk. This is the same frontier drawn against a meter: what one Artificial Analysis Intelligence Index task actually costs, measured rather than derived, with the option to re-price it at whatever route the market is offering today.
THE COST MAP
Anthropic · routed via Claude Opus 5.5
† CONDITIONAL RE-RATES
WHAT YOU PAY TO THINK
Reasoning tokens are output tokens. 9 models in the catalogue — every Gemini Flash and Flash-Lite generation, plus Gemini 2.5 Pro — publish a distinct reasoning rate, and all 9 of them set it equal to the model’s own output rate. So thinking bills at the output rate everywhere, and the light bar below is the only part of the answer you ever read. The 48 models with published per-task volumes are listed here; the rest have no token breakdown to split.
THE CACHE-WRITE PREMIUM
Eleven models charge a premium to write the cache rather than a discount: every Claude at exactly 1.25× the input rate, GPT-5.6 Sol at 1.25×, and Claude 3 Haiku at 1.20×. Below the break-even reuse count, turning caching on raises the input bill. Move the two controls above and watch the sign flip. Separately, only 61 of the 70 priced models publish a cache-read rate at all; for the rest the slider correctly does nothing.
No cache pricing published. GLM-4.7 Flash, GPT-3.5 Turbo, GPT-4, GPT-4 Turbo, gpt-oss 20B, Nemotron 3 Super, Qwen3.5 122B-A10B, Qwen3.5 27B, Qwen3.5 9B. Their input bills at the list rate no matter where the slider sits — an absence of data, not a rate of zero.
SHOPPING AROUND
OpenRouter’s headline number is the cheapest route it can find, which is often a third-party host rather than the model’s own vendor. That is a real saving you can take, but it is not the list price, and the cheapest-route mode above is quietly built on it. These 14 models diverge by more than 10% from the price Artificial Analysis used when it measured them.
COST PER TOKEN
First-party API list prices per million tokens, last updated Sep 5, 2026. OpenAI prices were refreshed on that date; other vendors retain the September 1 snapshot. On July 30 OpenAI cut GPT-5.6 Luna by 80% and Terra by 20%; the next day DeepSeek re-released V4 Flash — the model whose April weights sit on the local graph on the model map — with agent-focused post-training and the same $0.14-per-million input price.
A downloaded GGUF has no meter: once weights fit your Mac, the marginal cost of a token is electricity. These list prices are what local inference competes with — and July moved the floor closer to zero.
DEEPSEEK API CHANGELOG
What changed on the API side of the open-weight model this site tracks locally, from DeepSeek’s update log ↗.
READ THE FINE PRINT
The default x-axis is Artificial Analysis’ own published cost for one Intelligence Index task, applied as published. It already carries each vendor’s cache discounts. This page does not reconstruct it from prices and token counts — an earlier attempt to do that missed the published figure by 0.5× to 3×, because the only input-token figure available sits on a different weighting than the per-task output figure.
Reasoning tokens are output tokens. Nine models publish a separate reasoning rate and every one of them sets it equal to their own output rate, so no model here bills thinking cheaply. With reasoning running 60–90% of output on the heaviest models, the token volume — not the posted price — is what decides the bill.
OpenRouter publishes the cheapest available route, frequently a third-party host. 14 models diverge from the vendor price by more than 10%, the widest being DeepSeek V4 Flash at roughly 8×. The re-priced mode is an estimate built on that route and on a per-task input volume you choose; it is labelled as an estimate everywhere it appears.
Anthropic and OpenAI charge 1.20–1.25× the input rate to write the cache. Written once and read once, that is a surcharge, not a saving; it only turns into a discount past each model’s break-even reuse count. Six models also carry conditional re-rates — long-context thresholds or DeepSeek’s UTC off-peak windows — and the chart uses only their base rate, so they are flagged with †.
Per-task token volumes and cost figures © Artificial Analysis, used with attribution; API list prices from OpenRouter’s public catalogue. Pricing snapshot generated Sep 22, 2026; first-party vendor board updated Sep 5, 2026. A missing number here always means it could not be read from a source, never that it is zero.