AI models compared: price and performance

How to read this

Tokens are the small chunks of text, often parts of words, that models read and write. Prices are each lab's standard API rate in US dollars per million tokens, taken from the lab's own pages. Input is what you send the model; output is what it writes back. The Checked column gives the date we last compared the price with its source.

Scores come from independent leaderboards that test every model on the same tasks. We don't use the scores labs publish about their own models, because each lab tests in its own way. Coding is Vals.ai Terminal-Bench 4.0 (the share of 66 long, realistic tasks an AI agent completes in a computer terminal). Reasoning is LiveBench Reasoning (the average score out of 100 across LiveBench's reasoning tasks). Speed is Artificial Analysis (the median number of tokens a model writes per second).

Most of these models can be set to think for longer before answering, at a higher cost. Each row under a model is one of those settings, called a thinking level. Where a leaderboard's table shows no result for a level, the cell says "No score", and we never fill the gap from another level or another source. "Level not stated" means the source didn't say which level it used.

The context window is the most text, counted in tokens, that a model can handle in one request. "Open" in the Weights column means the lab publishes the trained model, known as its weights, for anyone to download and run. "Open base model" means the lab publishes the weights of the model this one is built on.

Claude Opus 5.5: Vals says 22 of Opus 5.5's 198 coding attempts were answered by older Claude models through a provider fallback. Counting those as failures lowers its score to 58.08%. Artificial Analysis also marks its Opus 5.5 speed results as including fallback.

GPT-6 Astra: Prompts with more than 272K input tokens are billed at twice the input and cache rates and 1.5 times the output rate for the full request.

Gemini 3.8 Flash: The price shown runs until December 31, 2026. From January 1, 2027, it is $1.50 in and $7.50 out.

Grok 4.7: Prompts over 200K tokens cost $4 in and $12 out.

Muse Spark 1.3: Meta's Contributor tier costs $0.10 in and $0.20 out if you let Meta train on your prompts and completions.

DeepSeek V4.1 Flash: The price shown is DeepSeek's peak-hours rate. Off-peak is half. Vals lists max effort in its model settings, though its write-up says its evaluations ran at high effort.

Mistral Large 4: Released as a public preview. The price shown is the list price. Mistral set launch pricing at 50% off for two weeks, which came to $0.68 in and $2.09 out when we checked. Mistral says open weights are coming soon.

Qwen3.8 Max: The price shown is Alibaba's international rate. Alibaba says Max is based on Qwen3.8-2.4T-A95B, whose weights it publishes. Artificial Analysis tested the September 2 snapshot of Max; the other two leaderboards don't say which snapshot they tested.

Kimi K3: Moonshot's changelog dates the release to July 2026 without a day.