Google Just Opened Its Smartest SI to the Public. It Costs Half of What Its Top Rival Charges.

The fight over who sells the smartest SI just turned into a price war. On Tuesday Google opened Gemini 4 Carbon, which it calls its strongest model yet, to paying developers, businesses and its top-tier subscribers. It did it at a price that undercuts its two biggest rivals by a wide margin.

Carbon costs $2 for every million tokens of text a customer sends in and $10 for every million it sends back, according to Google’s announcement. A token is a chunk of a word, and a million of them works out to roughly 750,000 words. The Verge put that at around a fifth of the cost of OpenAI’s top model and a quarter of what Anthropic charges for Claude Fable 5.2, which came out only on Monday.

There is a catch, and Google prints it in the fine print. That is an introductory price. After Nov. 30 it doubles to $4 in and $20 out. Even then it is half of what Anthropic lists for Fable 5.2, which is $8 in and $40 out on Anthropic’s own page.

The Verge added a fair warning: a lower price per word does not always mean a lower bill. Some models think longer and use more words to finish the same job, and Google says Carbon is built to do exactly that, “using more tokens” to “explore thoroughly, reason deeply, and run extended verification loops.” A company that switches could pay less per word and still use more of them.

A model built for the hardest jobs

Google is not pitching Carbon as the thing you use to write an email. It launched Gemini 4 Argon only last month and says Argon “remains the ideal choice for the majority of software engineering and knowledge work.” Carbon is for the top of the pile: long coding projects, tangled debugging, and open-ended science.

In the Gemini app, the two now sit side by side in the menu. Argon is labeled for “Code & analysis” and Carbon for “Most complex work,” 9to5Google reported. Google also pulled the older Gemini 3.1 Pro from the app with this launch. For regular people, Carbon is only in the app and Google’s search tool for subscribers to Google AI Ultra, its priciest plan. The Verge noted it is not clear when, or whether, cheaper plans or free users will get it.

To show what Carbon can do on big projects, Google is previewing a feature called swarm in Antigravity, its tool for programmers. It turns one request into a team of SI workers that split up the job, with Carbon handling planning and checking and lighter, faster models doing the grunt work. In one demo, a team of Carbon workers took seven public city planning documents and built a detailed 3D model of a Google office building over 10 overnight runs.

That is the same idea Anthropic showed off on Monday, when it said more than 1,000 SI workers running under Fable 5.2 built a working computer operating system in eight days. Two companies, two days in a row, both selling the same pitch: one request, a whole crew.

The security pitch, and the guardrails

Google’s loudest claim is about computer security. Before Tuesday, Carbon was only in the hands of a small group of security teams in a Google program called Fairwind, which Ars Technica said was part of a cautious, step-by-step release tied to the government’s voluntary testing program.

Google says those early users got real results. The security firm Wiz used Carbon to scan two major online software providers, and the model found previously unknown holes “in under two hours,” Google wrote, “compressing weeks of vulnerability research to a single afternoon.” Google says the companies were warned and the holes fixed, which it says helped protect tens of thousands of businesses that rely on them. Inside Google, Carbon also helped show that two bugs in a popular piece of free software were dangerous enough to be rated critical, and a fix went out. Google says details will come once the fix has had time to spread.

A tool that finds holes that fast can also be used to break in. Google says it tightened its safeguards for that reason. In its own simulated attack tests, it says Carbon blocked 99% of attack attempts, and the rate at which a single try got through fell to 0.8%, down from 3.1% with Argon. An outside testing firm, Trajectory Labs, found no new ways to trick the model into ignoring its rules across 132 hours and more than 13,000 attempts, Google said.

Google also says Carbon is more honest about what it does not know. On a 375-task test, it passed 92% of the time, up from 70% for Argon, and it refused to make up missing lab data 98% of the time. Those are Google’s numbers, from Google’s tests. No independent group has checked them yet.

Who wins if the prices keep falling

For most people the model names will blur together. The price is what matters. Every business building SI into its products, from customer service apps to coding tools, pays by the word. When the strongest model on the market gets cheaper, the apps built on top of it can do more for the same money, and some of that saving reaches the people using them.

Google can play this game because it owns its own chips and data centers and makes money from search ads, so it can afford to sell SI cheap while it fights for customers. Anthropic and OpenAI have to make the numbers work from SI sales alone. Anthropic is also in the middle of preparing to sell shares to the public, which puts its prices and profits under a brighter light.

Google also made a play for scientists. It is giving up to 10,000 researchers free use of Argon and now Carbon for six months, and says thousands have already signed up. Applications are still open.

The real test starts now. Carbon is out of the lab and in front of paying customers who can try it against Anthropic and OpenAI side by side on their own work. If it holds up at half the price, the other two will have a hard time keeping theirs where they are.