This page explains where every number on Polarison comes from, how we turn price sheets into cost estimates, and how to reuse our data. Prices were last verified on September 14, 2026; benchmark data was retrieved on September 15, 2026.
Prices
- Every price is copied by hand from the provider’s official pricing page — never from resellers, aggregators or press coverage: OpenAI, Anthropic, Google, xAI and DeepSeek.
- Each model stores the date its price was last checked, shown on its page and in the dataset.
- An automated job downloads every official pricing page once a day and flags any change in the text around prices, plus any price in our data that no longer appears on the page. Flagged changes are reviewed by hand before the data is updated, and notable ones are logged on the pricing changes page.
- We list the standard tier in USD per 1 million tokens. Discounts (batch, flex, off-peak) and surcharges (fast or priority modes, data residency) are described in each model’s notes rather than mixed into the main price.
- Temporary promotional prices are listed as the current price, with the end date and the future price in the notes.
Cost estimates
cost per request = (fresh input × input rate + cached input × cached rate + output × output rate) ÷ 1,000,000
- Monthly cost is the cost per request times the number of requests.
- When a provider publishes a long-context threshold (for example 200K prompt tokens), prompts above it use the higher rate.
- Rankings by price use a blended price of (3 × input + output) ÷ 4, which approximates typical chat and retrieval traffic.
- Estimates exclude cache-write surcharges, taxes, tool fees, and volume or negotiated discounts. Tokenizers differ between providers, so the same text can produce different token counts.
Benchmarks and capability
- Capability data comes from Epoch AI, an independent research organization, under the CC BY 4.0 license.
- The headline number is the Epoch Capabilities Index (ECI), which combines many benchmarks into one general-capability scale. We also show GPQA Diamond, FrontierMath (Tiers 1–3), SimpleQA Verified, ARC-AGI-2, DeepSWE, APEX-Agents.
- We only attach scores to an exact model match. If Epoch AI has not rated a model, it stays unrated instead of borrowing a sibling model’s score.
- Scores are the best result Epoch AI recorded for each model, usually at its highest reasoning setting.
- When two models’ ECI confidence intervals overlap, we call them roughly tied. Benchmark gaps under 2 points are treated as even.
- A model is marked best value when no cheaper model we track, by blended price, has a higher ECI.
Open data
The full dataset — prices, limits, notes and capability scores for every model — is free to download:
- models.json — structured JSON with sources and units
- models.csv — one row per model, for spreadsheets
- llms-full.txt — the same reference as Markdown, for AI assistants
Polarison’s compilation is licensed CC BY 4.0: you can reuse it, including commercially, as long as you credit the source. Please cite:
Polarison, “AI API pricing and capability dataset”, https://polarison.com/data/models.json. Benchmark data: Epoch AI, “Capabilities & Benchmarking”, epoch.ai/benchmarks (CC BY 4.0).
Corrections
Providers change prices without notice. If you spot an outdated or wrong number, please contact us with a link to the official page and we will correct it.
Independence
Polarison is not affiliated with any AI provider and is funded by advertising. Advertisers have no influence on prices, scores or rankings.