A woman is calculating expenses using a calculator and papers at her desk

Frontier AI model API prices: a dated tracker of what each lab charges

By Himanshu Sakre·

Photo: https://kaboompics.com/ / Pexels

Per-million-token prices for the models released in the last two weeks, with the date, the source and the catch for each. Updated through the correction loop as prices change.

List prices are the first thing buyers compare and the least reliable. This tracker records what each lab charged per million tokens for its newest models as of November 19, 2026, with the date, the source and what the number leaves out. It is a living page: when a price changes, we update it through a dated correction rather than writing a new article.

The table, newest first

Prices are US dollars per million tokens, input then output. "Reported" means we are relying on press coverage of the vendor's page, not on the page itself.

  • Gemini 4 Neon (Google, November 18, 2026): $3 input, $15 output, cached input $0.15. Context window 1 million tokens. Source: Google's launch post. Artificial Analysis lists the same $3 and $15 and a blended price of about $2.21.
  • GPT-6.2 Astra (OpenAI, November 17): $8 input and $40 output, reported by The Neuron, down 20 percent from GPT-6 Astra. Our launch report adds that long prompts above 272,000 input tokens cost more.
  • GPT-6.2 Sol (OpenAI, November 17): $2 input and $10 output, reported.
  • MAI Code 2 and Flash (Microsoft, November 17): available in GitHub Copilot on usage-based billing. Microsoft shows costs at a 50 percent promotional rate. No per-token price appears in the sources we opened.
  • Gemini 4 Carbon (Google, November 10): $2 input and $10 output introductory, then $4 and $20 after November 30, 2026. Source: Google's announcement.
  • Muse Ember (Meta, November 9): $0.10 input, $0.004 cached input and $0.50 output for prompts up to 200,000 tokens. Longer prompts cost $0.30, $0.012 and $1.50. Context window 1 million tokens with up to 512,000 tokens of output. Source: Meta's announcement.

What the table says

Three patterns stand out. First, the premium tier is not falling in price uniformly. OpenAI cut Astra's price, yet Google is doubling Carbon's after its introductory window. A buyer who locked in a Carbon budget in early November faces a 100 percent increase on November 30.

Second, the range is enormous. Ember's output price is $0.50 and Astra's is $40, an 80-fold gap. The gap is not only quality. Meta describes Ember as built for repeated agent tasks such as tool calls and data extraction. On Artificial Analysis's index, Meta reports Ember at 36, level with Claude Haiku 5.5 and above GPT-6 Luna's 28, while Haiku still led on FrontierCode 2.0. Cheap models win on volume work and lose on the hardest tasks.

Third, effort settings change cost. Artificial Analysis lists Neon at 61 at xhigh thinking and 44 at low, with a full index run costing about $6.44 at the top setting. A model's list price tells you the cost of a token, not of an answer.

Sleek laptop showcasing data analytics and graphs on the screen in a bright room
A spreadsheet on a laptop, the place most teams compare model prices. Photo: Lukas Blazek / Pexels

How to use the numbers

Use list price to eliminate, not to choose. If a model's output price is ten times your budget, drop it. If two models are within a factor of two, compare cost per accepted result on your own tasks. Our reading of Google's claim for Neon, that cost per task at high thinking is 32 percent below Carbon's, is a reminder that vendors now publish cost-per-task as well as price. That is a better metric, though it is still the vendor's own.

Our analysis of the Neon launch covers that claim in more detail.

A worked example

To see how the table changes a budget, take a job that uses 10 million input tokens and 2 million output tokens. At list price, ignoring caching and long-prompt surcharges, that costs about $60 on Gemini 4 Neon ($30 input plus $30 output), about $160 on GPT-6.2 Astra, about $40 on GPT-6.2 Sol, about $2 on Muse Ember, and about $40 on Gemini 4 Carbon at the introductory price, rising to about $80 after November 30. These are our calculations from the listed prices, and they assume each model needs the same number of tokens, which is rarely true. A model that thinks longer, or retries, can erase a price advantage.

The example also shows why the cheapest model is not automatically the right one. At about $2, Ember costs a thirtieth of Neon for the same token volume, yet Meta's own table shows it trailing on FrontierCode 2.0. If a cheap model fails a third of your tasks and a person must redo them, the savings shrink fast.

What this tracker does not cover

It does not include open-weight models served by third parties, whose prices vary by host. It does not include consumer subscriptions. It does not include batch or priority tiers, which change effective price. Hydrogen, Google's faster model, has been reported at $1.50 and $7.50 on a cloud price list, but Google declined to comment on availability when it was reported, so we leave it out until Google confirms.

Price is also only one line of a model's terms. Rate limits, regional availability, data retention and access tiers all differ, and several of these models, such as Carbon's cyber variant through the Fairwind Program, are restricted to vetted users. A cheaper model that you cannot call at your volume is not cheaper.

Corrections and updates

We will update this page when Google's Carbon price changes on November 30, when OpenAI's Ultrafast tier rises to six times standard pricing on March 1, 2027, and when Microsoft publishes a per-token price for MAI Code 2. Every change will carry a dated note. If you spot a price we have wrong, tell us and we will correct it publicly.

Our position: price is now a feature competitors move weekly. Treat any number here as a snapshot, and expect the next lab release to reset the table.

Frequently asked questions

What are the current frontier AI model API prices?

As of November 19, 2026: Gemini 4 Neon $3 and $15 per million tokens, GPT-6.2 Astra $8 and $40 (reported), GPT-6.2 Sol $2 and $10 (reported), Gemini 4 Carbon $2 and $10 until November 30, and Meta Muse Ember $0.10 and $0.50.

Is Gemini 4 Carbon getting more expensive?

Yes. Google lists $2 input and $10 output per million tokens as introductory pricing, rising to $4 and $20 after November 30, 2026 for the public Carbon model.

What is the cheapest AI model released this month?

Among models in this tracker, Meta's Muse Ember at $0.10 input and $0.50 output per million tokens for prompts up to 200,000 tokens, with a 1 million token context window.

Does list price show what a task costs?

No. Artificial Analysis shows Neon scoring 61 at xhigh thinking and 44 at low, with a full index run costing about $6.44 at the top setting, so effort and token use change the real cost.

How often is this tracker updated?

It is updated through dated corrections whenever a listed price changes, including Carbon's change after November 30, 2026 and Microsoft's per-token price for MAI Code 2 once published.

Sources

What each one is, and whose it is.

  1. 1

    Introducing Gemini 4 Neon, Google (November 18, 2026)

    Vendor announcement
  2. 2

    Gemini 4 Neon model page, Artificial Analysis (November 18, 2026)

    BenchmarkIndependent of the vendor
  3. 3

    Gemini 4 Carbon now available, Google (November 10, 2026)

    Vendor announcement
  4. 4

    Muse Ember, Meta (November 9, 2026)

    Vendor announcement
  5. Press reportIndependent of the vendor
  6. Documentation