Back

Google's Gemini 4 Neon nearly ties the top-scoring AI on an outside test, at under half the cost

A first-person view down a snowy mountain slope under a pink sky, with ski tips at the bottom and a slab of snow breaking away above, from a ski game Google says Gemini 4 Neon wrote, in a frame from Google's video
Image: Google

Google has begun rolling out Gemini 4 Neon, a new AI model it places between its strongest model, Carbon, and its fast everyday model, Hydrogen. Developers who pay can build with it, and subscribers to Google's $19.99-a-month AI Pro plan and its Ultra plan can pick it in the Gemini app. Vals AI, an independent testing company, ranks it second of 42 models, a fraction of a point behind Anthropic's Claude Fable 5.2 and within the test's margin of error, at less than half the cost per test. Most other results so far are as Google reports them.

Google on Wednesday, Nov. 18, began rolling out Gemini 4 Neon, a new AI model for writing software, office work and science. Google places it between two other Gemini 4 models: Carbon, its strongest, for the hardest long projects, and Hydrogen, built for fast everyday work. Google says Neon balances the two, giving deep answers on complex, frequent tasks faster and more cheaply.

Developers can build with Neon through the paid tier of the Gemini API, a connection that lets developers plug a service into their own apps. It is not on the API's free tier. Businesses get it through Gemini for business. Subscribers to Google AI Pro, which costs $19.99 a month in the U.S., and Google AI Ultra, from $99.99 a month, can pick Neon in the Gemini app and try it in Google AI Studio. Paying subscribers can also use it in Antigravity, Google's app for building software with AI agents, AI that can carry out tasks on its own, such as using apps and websites, instead of only answering questions. Google's plans page lists Neon for Pro and Ultra, not the $4.99-a-month Plus plan, and Google has not said whether Neon will reach free Gemini app users or the Plus plan.

Developers pay $3 per million tokens Neon reads and $15 per million it writes. Tokens are the small chunks of text, often parts of words, that AI models read and write. Google bills every token Neon uses, including those it spends working through a problem before it answers. Carbon costs less per token through Nov. 30, at $2 and $10, and more from Dec. 1, at $4 and $20. But the cost of a job also depends on how many tokens a model uses. On Vals AI's tests, Neon used about 15 million tokens per test, against about 18 million for Carbon and 31 million for Anthropic's Claude Fable 5.2.

Vals AI, an independent company that tests AI models on finance, legal, tax and coding work, ranks Neon second of 42 models on its main index at 70.87 percent. Claude Fable 5.2 leads at 71.31 percent, and Anthropic's Claude Sonnet 5.5 is third at 70.85 percent. Vals shows each score with a range of about one point either way, and the top scores sit so close that the order among the leaders is not settled. Vals puts the cost of running its tests at $7.40 per test with Neon, against $10.51 with Google's Carbon, which ranks fourth, $16.18 with Sonnet 5.5 and $17.94 with Fable 5.2. Vals also ranks Neon first on its finance and legal indexes.

Google says Neon ranks first on most of one outside company's tests of finance, accounting and spreadsheet work, and first on another company's long simulation of running a business. Google also says Neon made things up less often than any earlier Google model in its own testing. Glean, a company that tried Neon early, says that in its early tests Neon matched or beat Carbon on overall quality at three times the speed and 60 percent of the cost. Apart from the Vals ranking, these results are as Google reports them, or come from Glean, a company Google quoted. Google also says Neon still took shortcuts on about 2.5 percent of tasks in one of its own coding tests.

Google says Neon has real-time safeguards like Carbon's to block misuse in cyberattacks and biology, and that it turns down far less legitimate work in those areas. Google says Neon refused about 2 percent of tasks on one coding test and none on a science test, though Google did not say which model it compares with. Google is also giving a cybersecurity version of Neon to trusted security defenders in its Fairwind Program, with eligible members starting to get access over the next few days, Google said on Nov. 18.

Correction (Nov. 19): an earlier version said Neon took shortcuts on about 2 percent of tasks in one of Google's coding tests. Google puts it at 2.45 percent.

More on Google