AI Pricing Calculator

How to use this calculator

Input tokens = everything the model reads (your prompt, attached files, conversation history). Output tokens = what the model writes back.

Rule of thumb: 1 token ≈ 0.75 words. A 1,000-word document is ~1,300 tokens.

Cached input ratio matters because most providers offer steep discounts on tokens that hit their cache — Anthropic reads cache at a fortieth of the fresh-input price on Fable 5.1, OpenAI at a tenth on GPT-6 Astra. For applications with consistent system prompts or repeated context, cache hit rates of 50–90% are achievable.

How to read the results

The list is sorted cheapest first and the top card is marked best value, but that is best value for the exact tokens you entered. The calculator cannot know how many tokens a model will actually spend on your task, and that is where the real gap between models lies. Our GPT-6 Astra reviewis the case in point: Astra and Claude Fable 5.1 carry the identical $10/$50 list price and will show identical bills here, yet on Artificial Analysis's measured runs Fable 5.1 spent 2.7× more to finish the same benchmark because it emitted three times the output tokens. When two models are close on this page, use the cost-per-task column on the leaderboardto break the tie, and treat reasoning-heavy “max effort” settings as an output multiplier you are choosing to pay for.

Read the tier badges as a hint about behaviour, not quality. Frontier models write longer, think longer and cost more per token; Fast & Cheapmodels are the right default for high-volume classification, extraction and routing where a frontier model's judgement is wasted. Reasoningmodels bill their hidden thinking as output tokens, so a small visible answer can carry a large output bill. The notes under each card record repricings — GPT-5.6 Sol went to a promotional $4/$20 on 21 August and GPT-6 Sol then halved it to $2/$10, Sonnet 5's $2/$10 launch price was made permanent instead of rising to $3/$15, Claude Opus 5.5 arrived a fifth under Opus 5 at $4/$20 — because the date a price changed is as useful as the price.

Where the prices come from

Every figure is a provider's published API list price in US dollars per million tokens. The base table was verified against the official pricing pages in May 2026; the September 2026 frontier launches were added from published launch pricing, last on September 23, 2026. Sources: Anthropic, OpenAI, Google, xAI; other providers as linked from their tool pages. Providers update prices regularly — check the official page before committing to a production budget. Open-weight models like Llama 4 405B and DeepSeek V4 can also be self-hosted for unlimited usage if you have the hardware, which this calculator does not model.

What is not included: batch-API discounts (typically 50% in both directions), long-context surcharges (GPT-6 Astra bills prompts above 272K tokens at 2× input and 1.5× output), cache write fees where a provider charges them, image or audio tokens, and subscription plans. For subscriptions the question is different — whether a $20 seat pays for itself — and our Claude and ChatGPT Plus reviews answer it; for metered coding tools, the AI-credit explainer shows how a credit balance converts to tokens at these same rates.

Frequently Asked Questions