AI Model Comparison

Model index

AI Model Index: Every Model We Track, with Price and Context

Ten LLMs compared on price per million tokens, context window and licensing — checked 22 November 2026.

Standard API rates · USD per 1M tokens · checked 2026-11-22

Ten models, one table

Every figure below was read from the provider's own model listing or a public model catalogue on 22 November 2026. Prices are USD per million tokens at the standard (non-batch) rate. If a provider publishes a batch, cached-input or promotional rate, that is noted under the table rather than mixed into the main numbers.

Model Maker Context Input $/1M Output $/1M Open weights
GPT-6.2 Astra OpenAI 1.05M 8.00 40.00 No
GPT-6.2 Astra Pro OpenAI 1.05M 8.00 40.00 No
Claude Fable 5.2 Anthropic 1M 8.00 40.00 No
Claude Opus 5.5 Anthropic 1M 4.00 20.00 No
Gemini 4 Neon Google 1.05M 3.00 15.00 No
Gemini 4 Hydrogen Google 1.05M 1.50 7.50 No
Qwen4-Max Alibaba 1M 1.50 4.50 No
Mistral Large 4 Mistral 1M 0.70 2.10 Yes (preview)
DaisyRose 27B RadixArk 262K 0.15 1.875 Yes (Apache 2.0)
Inkling 2 Mini Thinking Machines 1.05M 0.15 0.50 Yes

Notes: these are standard rates. GPT-6.2 Astra is listed at $4.00 input / $20.00 output for batch work and Gemini 4 Neon at $1.50 / $7.50, both returning results later for roughly half price. Mistral Large 4's $0.70 / $2.10 is a preview promotion; Mistral's list rate is $1.36 / $4.18. Cached input is cheaper again where the provider supports it — Inkling 2 Mini lists $0.015 per million cached input tokens, Mistral Large 4 $0.07, and Qwen4-Max and Gemini 4 Hydrogen $0.075. Batch, cache and promotional rates are not comparable with the standard column, so we keep them here rather than in the table.

One model that is not in the table yet: Gestalt, the unnamed stealth model that has led a new design leaderboard since 18 November 2026. It has no public access and no published price, so there is nothing to list — the tracker page follows it until there is.

The closed frontier models

GPT-6.2 Astra — OpenAI

The top of OpenAI's price list, at $8 input / $40 output per million tokens, with a context window of just over a million tokens and image input. Astra Pro is listed at the same headline price, so the difference between the two is not cost — it is which one the provider routes the hardest requests to. If you are comparing them, compare on output quality for your task, not on price.

Claude Fable 5.2 — Anthropic

Anthropic's most capable model, released 9 November 2026 at $8 input / $40 output, with a 1M context window, 128K max output and always-on thinking. Fast mode doubles the rate to $16 / $80 for up to 2.5× faster output; cache reads cost $0.20. Anthropic positions it for the hardest technical work — long-horizon agents, days-long problems, self-verifying research loops — and on per-task cost it has beaten far pricier models on coding benchmarks. The catch is the price: at twice Opus 5.5's rate, you pay it only when Opus at high effort falls short.

Claude Opus 5.5 — Anthropic

The model Anthropic recommends for most tasks, released 22 September 2026 at $4 input / $20 output, with a 1M context window and always-on thinking (default effort medium; Fable 5.2 defaults to high). It sits between Gemini 4 Neon ($3 / $15) and the $8 / $40 tier, at the same 1M context as both. If your instinct is "I'll just use Anthropic's flagship", this is the one — Fable is the escalation, not the default.

Gemini 4 Neon — Google

The cheapest of the frontier three at $3 input / $15 output, with the same 1.05M context as Astra. At these rates one million output tokens costs $15 on Neon and $40 on Astra — the biggest price gap among the closed models in the table. Google's listing positions Neon for agentic coding and long output. The Neon page tracks the price ladder against Carbon's promo rates and the AA Index scores.

Gemini 4 Hydrogen — Google

The newest Google row in the table, made generally available on 12 November 2026: $1.50 input / $7.50 output, flat across the full 1,048,576-token window, with text, image, video, audio and PDF input. Cached input drops to $0.075 — a 95% implicit-cache discount — and there is a rate-limited free tier. Output is capped at 65,536 tokens by default, extending to the full window with long decode. At half of Neon's output price for the same window, it is the cheapest way to put a million tokens of multimodal context in front of a Google model. See the Hydrogen page for the full breakdown.

Qwen4-Max — Alibaba

Released on 16 November 2026: a 2.4-trillion-parameter sparse mixture-of-experts backbone with about 71B parameters active per token, native image input and a 1M-token context window, listed at $1.50 input / $4.50 output. The output price is the lowest of any closed model in this table — 60% of Gemini 4 Hydrogen's $7.50 — and Alibaba publishes average cost per Work360 task falling from $4.43 (Qwen3.8-Max) to $1.67. Weights are not released for this model, unlike the open-weight Qwen3.8 checkpoints. The Qwen4-Max page has the benchmarks and the price ladder.

The open-weight models

Mistral Large 4 — Mistral

Mistral's open-weight flagship, in public preview since 6 October 2026 at a promotional $0.70 input / $2.10 output (list rate $1.36 / $4.18), with cached input at $0.07. It is a 1.05-trillion-parameter sparse MoE with 52B active parameters per token, natively multimodal — text and image in, with a 1.6B-parameter vision encoder — over a 1M-token window, and the weights shipped at the end of October. It is the cheapest closed-quality multimodal option in the table, with the caveat that a preview price is not a promise: budget against the list rate.

DaisyRose 27B — RadixArk

$0.15 / $1.875, Apache 2.0, 262K context. This is the middle ground between a hosted frontier API and a tiny utility model: cheap enough to run at volume, big enough to hold a multi-step coding task. Because the weights are open you can also self-host it and pay for GPUs instead of tokens.

Inkling 2 Mini — Thinking Machines

$0.15 / $0.50 with a full 1.05M context window, open weights, and 3.4B active parameters out of 26.4B total in a mixture-of-experts design. It is the cheapest way in the table to put a million tokens of context in front of a model — about 3.75 times cheaper on output than the next cheapest model here. At $0.50 per million output tokens you can afford to run it on every request and keep the expensive model for the cases that fail. The Inkling 2 Mini page has the benchmarks, the Ollama setup and the local-hardware requirements.

What the table does not tell you

Price and context are the two numbers that decide your bill, so they are the two we lead with. They are not a quality ranking. No column here says which model writes better code, argues better or hallucinates less — for that you need benchmark scores and, more importantly, your own evaluation on your own data. The benchmark guide explains what the public leaderboards actually measure and where they mislead.

How to use this page

  1. Set a ceiling on cost. Take your monthly output tokens and multiply by the output price. That number, not the input price, is usually what surprises people.
  2. Check the context you actually need. If your longest prompt is 80K tokens, a 1.05M window is a nice-to-have, not a requirement — and you should not pay frontier prices for it.
  3. Decide whether you need open weights. Open weights mean you can self-host, fine-tune and keep data inside your own network. A model you cannot self-host is a vendor dependency, and that belongs in the decision.
  4. Confirm on the provider's page. Every price here has a date on it. Check that date against today before you budget.