AI Momentum 107.6 +2.7%GitHub Pulse 5172 -5.9%Capital Flow $204.6B +5804.2%Hiring Heatmap 0.0 -100.0%
Advertisement
Models & Research

Google's new mid-priced AI model nearly matches its flagship at less than half the cost per job. The saving mostly starts in December

Google released Gemini 4 Neon on November 18, its fourth Gemini 4 model in about seven weeks. It is pitched as a middle option: close to the quality of Google's flagship, Gemini 4 Carbon, but faster and cheaper for everyday business work.

For buyers, the headline numbers hold up better than most launch claims do. Artificial Analysis, an independent firm that tests AI models, found Neon roughly level with Carbon on its main quality index. At Neon's default setting, it put the cost of each test job at less than half of Carbon's.

There are two catches. The saving is measured against Carbon's full price, which does not apply until December 1. And Google's own scorecard compares Neon only with rival models it beats, leaving out the two that score higher.

What Neon costs, and when the math changes

Neon costs $3 per million words of input and $15 per million of output, measured in tokens, the small text chunks AI companies bill by. Google's AI Pro and Ultra subscribers get it in the Gemini app, where it replaces the older Argon model. Developers and businesses can buy it through Google's paid API and its enterprise AI platform.

Carbon currently sells for $2 and $10 under an introductory discount that ends November 30. It then doubles to $4 and $20. So until November 30, Neon is 50% more expensive per token than Carbon. From December 1, it is 25% cheaper.

Per-token prices only tell part of the story. Models differ in how much they "think" before answering, and that extra output is billed too. That is why independent testers increasingly report cost per task, the total bill to finish a standard job.

On that measure, Artificial Analysis said Neon at its default high setting matches Carbon at less than half the cost per task, using Carbon's full December price. Its published figures put that at about $3.74 per task for Neon versus $7.76 for Carbon. By machine.net's own calculation, halving Carbon's figure to reflect the current discount puts the two at roughly the same cost per task today, about $3.74 against $3.88.

Google's own claim is more modest. Its launch post for Antigravity, its coding tool, says Neon costs 32% less per task than Carbon at the default setting. The post gives no breakdown behind that figure.

How it stacks up against OpenAI and Anthropic

Google's model card, its official spec sheet, compares Neon with Anthropic's Claude Opus 5.5 and OpenAI's GPT-6.2 Sol. Neon leads on most of the 19 tests listed. Opus 5.5 wins two coding tests, and Sol wins one on reading charts.

Missing from that table are OpenAI's GPT-6.2 Astra and Anthropic's Claude Fable 5.2, each company's top model. Artificial Analysis ranks Neon third overall on its index, behind both. Neon's best score there, 61, comes from its highest "extra high" setting. At the default setting it scores 60, level with Carbon.

Neon looks strongest at agent-style work, where an AI carries out a multi-step job on its own. On Artificial Analysis's coding-agent test, Neon scored 59, a point above both Opus 5.5 and Fable 5.2. Its average cost per task was $6.86, versus $12.61 for Opus 5.5 and $15.47 for Fable 5.2. The testers note the comparison runs each model in its own maker's coding tool, so it is not a perfect like-for-like test.

Neon also topped AutomationBench, a business-workflow test from Zapier, at 56.3% in Google's figures. Opus 5.5 scored 42.5% and Sol 41.3%, by the same account.

What buyers should watch

Neon produces text quickly. Artificial Analysis measured it at about 110 tokens a second, nearly double Carbon's speed. But it also produces about 15% more output per task than Carbon, so its lower bill comes from its price, not from doing the same work with less.

Neon also took about 21 minutes per index task at its top setting. That beats Carbon's roughly 36 minutes. But the review site Fello AI, citing Artificial Analysis data, notes that is well above the typical figure of about 12 minutes for comparable models.

Early hands-on tests are small but encouraging. A DataCamp reviewer asked Neon and Carbon to build a working chess program three times each. Neon succeeded all three times at about 95 cents a run. Carbon succeeded once, failing twice when it hit an output limit.

Google has not yet published its full safety report for the model, saying it will come in the coming days. Almost all benchmark scores in the launch materials are Google's own, so they have not been independently checked.

The bigger signal is pace. Google has shipped four Gemini 4 models in under two months. For companies signing AI contracts, that argues for short commitments and testing on their own work before switching. Today's price advantage may not survive the next release.