Google Opens Up Carbon: 60 Points vs. 65 for Fable 5.2, at a Price 4x Lower

Google has opened Gemini 4 Carbon to everyone — the day after Fable 5.2. It sounds like an arms race, but for practitioners something else matters more: Carbon is close to the leader in quality and noticeably cheaper on the price list. There is, however, a caveat that Google itself states openly.

Where to Get It and What It Costs

Access is already open through several channels. Paying customers get the API in AI Studio and Gemini for Business, while AI Ultra subscribers get the model in Antigravity, the Gemini app and AI Mode. Scientists get six months free: the program is designed for 10,000 people.

The launch price is $2 per million input tokens and $10 per million output tokens (an introductory rate, same as Argon). For comparison: Fable 5.2 costs $8 / $40, GPT-6 Astra — $10 / $50. By list price, Carbon is 4–5 times cheaper.

What's Inside and How It Stacks Up Against Competitors

The main feature for developers is the /swarm command in Antigravity. You write a single prompt, and the system assembles a team of agents from it. Carbon plays the architect and auditor: it breaks the task down and checks the result. Lighter models write code in parallel, each in its own isolated environment. Separate verifier agents check every change. The idea isn't new, and in regular mode such a setup will simply devour tokens, so it should be enabled deliberately.

Now the numbers:

• overall AA index: Fable 5.2 — 65, Carbon — 60, Opus 5.5 — 59, GPT-6 Astra — 57;
• FrontierCode 2.0: Fable 49.1%, Astra 42.8%, Carbon 41.1%;
• OSWorld 2.1 (computer use): Fable 89.3%, Astra 79.5%, Carbon 81.5%;
• AutomationBench: Carbon 53.9%, Fable 51.4%, Astra 41.4%.

Carbon is first on CursorBench 4.0, WANDR and AutomationBench, and costs more than half as much less per task than the next external model. On the new MKA benchmark (agents that work with documents containing drawings, charts and tables) it is second, but six times cheaper than the leader.

There are some striking examples too. The Wiz team used the model to find a 0-day at two large SaaS providers in less than two hours. In science, Carbon helped formally prove properties of the Signal protocol code: millions of lines in Lean in weeks instead of years.

The Catch: A Cheap Token Is Not a Cheap Task

Google itself doesn't recommend rushing to Carbon for everything. For most code and document tasks, Argon remains the "ideal choice." Carbon is meant for tasks at the limit: it researches longer, thinks deeper and runs verification loops. Hence the paradox: on the AA index, one task costs $7.76 with Carbon versus $6.24 with Astra, even though the token itself is 2.5 times cheaper. It simply uses noticeably more tokens.

A hypothetical example. Renaming a field across a hundred files or extracting a table from an ordinary PDF is overkill for Carbon: you'd pay for reasoning you don't need. But finding a vulnerability, designing a refactor across dozens of services or checking a complex piece of code — there the extra verification loops can pay off.

What to Do in Practice

• Don't switch everything at once. Run 20–30 of your hardest tasks on Carbon and compare the cost per task, not the price per token.
• If your pipelines are still on Gemini 3.1 Pro, moving to Carbon or Argon looks reasonable. The author of the AI Product channel would do exactly that.
• Enable /swarm selectively: parallel agents and verifiers burn through a budget quickly.
• Google's models remain strong in multilingual work, visual understanding, translation, document recognition, audio and video, and prices haven't been jacked up yet.

The limitations are also obvious. In raw intelligence Fable 5.2 is ahead (65 vs. 60 on AA), and on FrontierCode the gap is notable: 49.1% vs. 41.1%. Most of the figures come from Google's announcement and third-party indexes, and results on your tasks may differ. So the bottom line: Carbon is "almost top-tier for less money," but only if the task is truly hard. Now we wait for OpenAI's response.