Strategy5 minNewsroom

Google launches Gemini 4 Neon at $3 per million tokens

Gemini 4’s mid-tier model outperforms Carbon on CursorBench.

Vista diurna do campus Google's Bay View em Mountain View, com telhados em forma de tenda e painéis solares.

Google launched the Gemini 4 Neon on Wednesday (18th), a mid-tier model from the Gemini 4 family, with a list price of $3 per million input tokens and $15 per million output tokens, featuring a context window of 1 million tokens. The Neon sits between Hydrogen, the fast version for everyday use, and Carbon, which Google reserves for longer and more complex projects. According to the company's announcement, the model is now available on the Gemini API in Google AI Studio, Gemini for enterprises, the Antigravity development environment, and the Gemini application for paid subscribers of the Pro and Ultra plans.

This is the fourth new Gemini model in about a month, and the launch came a day after OpenAI introduced GPT-6.2 Astra and GPT-6.2 Sol. In the comparison chart released by Google, Neon competes against Claude Opus 5.5 and GPT-6.2 Sol, a comparison that, being from the provider itself, uses the tests and configurations it has chosen.

The key number is cost per task

The most useful independent data so far comes from Cursor, which publishes the CursorBench with the models running inside its programming agent. In version 4.1, Neon scored 63.7% on the standard effort (High), compared to 61.5% for the Gemini 4 Carbon at the same level and 57.3% for Hydrogen. Cursor estimates Neon’s cost at $6.54 per task, 23% lower than Carbon High's $8.52. Cursor itself notes that small differences in scoring may not be statistically significant.

The token comparison misleads this week. Carbon is at an introductory price of $2 and $10 per million tokens until November 30, then rising to $4 and $20 on December 1, according to the pricing table from Google reproduced by various outlets. Until then, Neon costs more per token than Carbon. The advantage measured by Cursor comes from Neon completing tasks using fewer tokens, not from a lower fee. Starting in December, Neon will also become cheaper per token, about 25% below Carbon.

Why the mid-tier model became the battleground

For corporate buyers, the launch underscores a shift in criteria. For two years, model selection started at the top of the table: the most capable model, then the price was negotiated. With Google, OpenAI, and Anthropic releasing mid-tier models that tie or outperform the top in specific tasks, the relevant metric shifts to cost per task solved, which depends on how many tokens the model uses for reasoning. A model with a lower fee that requires twice the reasoning ends up being more expensive.

This has direct implications for those budgeting agent projects. Consulting and managed services contracts that price AI by token volume are exposed to pricing changes like those of Carbon, which doubles in price on a fixed date. Contracts that price by delivered task shift this risk to the supplier, who now has an incentive to switch models whenever the balance of accuracy and cost changes.

The outlook in different markets

In the United States, the first effect appears in programming tools: Cursor is already running Neon and publishing the cost per task, and engineering teams that pay by usage will compare their December bills with their November bills, when Carbon doubles in price.

In India, where TCS, Infosys, Wipro, and HCLTech concentrate a significant portion of the world’s outsourced software engineering, a mid-tier model that scores above the top in coding tasks alters the productivity metric used to renegotiate application maintenance contracts and increases client pressure for fixed pricing per delivery.

In Brazil, companies still piloting code assistants on small projects face the same decision with currency cost: a dollar table that doubles on a fixed date weighs more when the bill is converted to reais. Predictability, more than the list price, tends to determine supplier choice.

Google has not disclosed how it plans to position Carbon after the adjustment. If Neon delivers more in coding for less, the question that clients will bring to the next renewal is what exactly the more expensive model in the family is for.

Sources

  1. blog.googlehttps://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon
  2. digitaltrends.comhttps://www.digitaltrends.com/computing/google-says-gemini-4-neon-can-hang-with-claude-opus-5-5-and-gpt-6-2-sol-for-less/
  3. the-decoder.comhttps://the-decoder.com/google-ships-gemini-4-neon-its-fourth-new-gemini-model-in-a-month
  4. alphasignal.aihttps://alphasignal.ai/news/google-s-gemini-4-neon-beats-carbon-on-coding-tasks-while-costing-23-less
  5. implicator.aihttps://www.implicator.ai/google-gemini-neon-task-cost/
  6. braindetox.krhttps://braindetox.kr/en/posts/gemini_4_carbon_general_availability.html

The week's analysis, by email

One weekly edition with what matters to people who decide. No ads, no sponsorship.

One-click cancellation, at any time.

Strategy →