Brief IA
›

Gemini 4 Neon Reduces Task Costs by 32%

AI Tools·Tom Levy·

Gemini 4 Neon Reduces Task Costs by 32%

Gemini 4 Neon Reduces Task Costs by 32%
Key Takeaways
1Google unveiled Gemini 4 Neon on November 18, 2026, its fourth model in seven weeks.
2Google announces $3 per million incoming tokens and $15 for outgoing tokens for its API.
3The model's documentation acknowledges that Neon can still generate hallucinations.
Why it matters : Google presents Neon as a compromise between speed and advanced capabilities for software, documents, and agents.

Google unveiled Gemini 4 Neon on November 18, 2026, its fourth model in this range within seven weeks. The company announces a task cost that is 32% lower than that of Carbon in a "high" setting evaluation scenario. This comparison does not correspond to a general reduction in rates.

Neon Costs More Per Token During Promotions

For the API, Google announces $3 per million incoming tokens and $15 for outgoing tokens. Until November 30, the promotional prices for Argon and Carbon remain lower: $2 and $10, respectively. The transition to the standard rates for these models, set to take effect on December 1, will raise them to $4 and $20.

Thus, the expense associated with a task does not equate to the token rate. The claimed advantage for Neon depends on the comparison made by Google and the reasoning level chosen, not on a savings applicable to all uses.

Use Cases for Agents, with a Risk of Hallucination

Google specifically targets Neon for software development, document processing, and agents that alternate between action, verification, and retrying. An agent can examine a software repository, investigate the cause of a bug, suggest modifications, and then execute tests. The model can also handle large tables or PDFs; its context window reaches one million tokens.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Four reasoning settings allow for adjustments in reflection time and expenses. However, the model's documentation acknowledges that Neon can still produce hallucinations.

Neon Scores and Other Gemini 4 Models

On Terminal-Bench 4.1, Google DeepMind assigns 65.8% to Claude Opus 5.5, compared to 65.2% for Neon. Google also reports a score of 46.2% for its model on FrontierCode 2.0 with the "high" setting. On the Artificial Analysis side, Neon achieves 58.6 points and ranks first in the Coding Agent Index.

On the Artificial Analysis Intelligence Index, its reported score is 61 points, placing it third with "xhigh," its maximum reasoning setting. The ranking in this composite index depends on the configuration. Google broadly presents Neon as an offering that combines advanced capabilities, speed, and efficiency in repetitive tasks, without the price of a high-end model.

The other models in the range have distinct priorities: unveiled on September 30, Argon targets complex reasoning, while Carbon, released at the end of October, focuses on difficult problems and cybersecurity. Introduced on November 12, Hydrogen aims more at speed and daily needs. Neon follows just six days later.

Access for Subscribers and Developers

Since November 18, the rollout in the Gemini application concerns Google AI Pro and Ultra subscribers. These paid offerings also provide access to Neon in Google Antigravity.

For developers, access is through the Gemini API within Google AI Studio. Companies also have specific access pathways.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.