Every major model, side by side
The products people use (ChatGPT, Claude, Gemini, Copilot, Grok, Meta AI, DeepSeek) are built on underlying models that change every few weeks. This page lists the models each developer currently makes available, with release dates, whether the weights are open, how much text they can take in at once, what they cost per million tokens through the developer’s own interface, and which benchmark claims they make. Every row links to the developer’s own documentation or announcement. Where a number comes from a company about its own model, the table says so.
How to read this. A token is about three-quarters of a word; prices are U.S. dollars per one million tokens in and out. Context is how much the model can hold in mind at once (1M tokens is several long novels). Open weights means anyone can download and run the model; closed means you can only use it through the company. Benchmarks in parentheses say who reported them; a lab’s number on its own model is a claim, not a measurement. See how models are tested.
What a million tokens costs
Output price, US$ per million tokens, current flagship and budget tiers (developer price lists)
Prices as listed on each developer’s pricing page in Sept 2026; they change often and vary with caching, batch discounts and context length. Open-weight models are shown at the developer’s own hosted price where one exists.
What this table cannot tell you
Corrections and additions: email with the developer’s documentation link. Changes are logged on the change log.
Who releases their weights
The single biggest fault line in the industry. Open weights let anyone run, study and modify a model; closed weights keep control (and the safety decisions) with the developer.
Current models by developer and openness
Count of rows in this table marked current
The picture inverted between 2024 and 2026. Meta, which built its AI reputation on the open Llama family, shipped its first closed model in April 2026. Chinese developers went the other way: DeepSeek, Alibaba’s Qwen team, Moonshot, Zhipu, MiniMax and Tencent now release most of their frontier models under permissive licenses such as MIT and Apache 2.0, and by mid-2026 the largest open-weight model in the world came from Moonshot. OpenAI released its first open-weight models since GPT-2 in August 2025; Google switched its Gemma family to Apache 2.0 in April 2026; NVIDIA and Mistral publish open models as a strategy. Anthropic, xAI, Apple, Microsoft’s MAI line and the Gemini and GPT flagships remain closed.
Why it matters for policy: the U.K. AI Security Institute found open-weight models trailing closed frontier models by four to eight months on the tasks it tests, which means anything a closed model can do today, anyone can run on their own hardware within a year. See risks and policy.