Brief IA
›

Google Launches Gemini 4 Neon, Priced at $3 per Million Tokens

Research·Tom Levy·

Google Launches Gemini 4 Neon, Priced at $3 per Million Tokens

Google Launches Gemini 4 Neon, Priced at $3 per Million Tokens
Key Takeaways
1Google begins the rollout of Gemini 4 Neon in its APIs, professional offerings, and subscription services.
2The pricing is $3 per million incoming tokens and $15 per million outgoing tokens.
3Google reports 39.0% on SWE-sweep in Antigravity, compared to 12.8% with mini-SWE.
Why it matters : teams have four levels of reasoning to adjust the capabilities utilized and their token expenditure.

Google begins today the rollout of Gemini 4 Neon, priced at $3 per million incoming tokens. The model is available in its developer tools, enterprise offerings, and subscription services. Google highlights long programming and analysis tasks, showcasing cost savings and protections for which it publishes evaluation results.

Security and Behavior: Announced Results

On Terminal-Bench 4.0, Google reports a cheating rate of 0.0% for Neon, the same as for Carbon. Another internal programming evaluation notes shortcuts in 2.45% of tasks, the minimum for the Gemini 4 family. The company claims to have trained the model to respect user limits, without skipping steps or fabricating results.

During internal programming uses, the thought and action misalignment monitor recorded lower alert rates than those of Carbon, according to Google. The company sees this as an indication of better alignment. It asserts that it retains the protections of Gemini 4 Carbon against serious abuses and instruction injections, with comparable real-time defenses against cyber abuse and biosafety risks, but significantly fewer refusals of legitimate tasks.

For secure code generation, Google places Neon ahead of Carbon in every difficulty category of AutoBaxBench, with a claimed record of 76% in the most difficult category. On WebExploitBench v1.2, it reports 51.26%, presented as a leading result achieved at half the cost of competitors. Google also claims top-tier performance for detection, validation, and correction of vulnerabilities, with greater speed and a better cost-effectiveness ratio than leading models.

Pricing, Access, and Reasoning Levels

Cached incoming tokens are charged at $0.15 per million, a 95% discount; outputs cost $15 per million. The context can accommodate 1 million tokens. Access is available through the Gemini API in Google AI Studio, as well as in Gemini for enterprises, via API or from the model selector. Google AI Pro and Ultra subscriptions provide access to deployment in Antigravity; the Pro and Ultra plans are also included in the Gemini application.

Google offers four reasoning levels to tailor the capabilities deployed and token consumption. The company claims to refine the progression with each version and asserts a reduced cost per task in third-party evaluations. At the high level, set by default, it reports a high score on the Artificial Analysis Intelligence Index with a per-task expenditure 32% lower than Carbon. On Vals and Agent Arena, it positions this same setting on the Pareto frontier, with costs approximately 50% to 80% lower than those of the best competing models.

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

The xhigh setting is aimed at difficult engineering, science, and analysis problems requiring more reasoning steps. Google assigns it the best intelligence score for Neon, with a per-task cost 36% lower than that of competitors configured at their maximum level.

Programming: Scores Vary by Environment

On SWE-sweep, Google reports 12.8% with mini-SWE, an execution environment where it presents Neon as the leader. The result reaches 39.0% in Antigravity, a claimed record and more than tripled score. Google attributes this gap to the Antigravity environment. It describes a model capable of debugging or reworking multiple files from a repository by making only targeted changes.

Neon can, according to the company, work independently on long tasks involving many tools or coordinate Hydrogen sub-agents in Antigravity. Google credits it with fewer unnecessary file reviews, progress reports, and reliable context compaction management. Its internal trials in Antigravity reportedly yielded its lowest rate of fabricated results, with good tracking of successive instructions. For considering user corrections during a task, Neon is presented as second in the new Agent Arena Index.

In the programming evaluations cited by Google, Neon ranks first in the Artificial Analysis Coding Agent Index, with a score of 58.6. It also achieves 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0, the latter result measured at the default high setting.

Results in Business and Science

Google also targets Neon for intellectual and scientific work, claiming rapid execution. The model can, according to the company, aggregate spreadsheets, PDFs, graphs, and audio or video content to produce structured deliverables. In the Gemini application, Google presents it as suitable for in-depth analyses, with quick, organized, and clearly written responses.

Among the independent evaluations it cites, Google reports a third place in the Artificial Analysis Intelligence Index: Neon scores 61 in xhigh, behind models described as larger and slower. On the Vals Index, its score of 70.9% earns it second place and a spot on the Pareto frontier. It ranks first on Legal Research Bench and Tax Agent Bench. On Surge AI, Google also claims first place in most tests of finance, accounting, spreadsheet modeling, and compliance with complex constraints.

For reading dense professional graphs, Chartography assigns it 73.5%, while Agents' Last Exam v1.1 gives 75.8%, presented as comparable to the best models. Google announces a first place on Zapier’s AutomationBench with 56.3%, as well as on Andon Labs' Vending-Bench 2, an extended business management simulation, with $16,843. On the new Agent Arena Index from Arena, Neon ranks second globally with 1336, from over 4,200 real agent sessions.

In scientific and mathematical reasoning, Google presents it as the best in its category on Terminal-Bench Science, at 70.5%. It claims the same status on RiemannBench with 84.0% and HLE-Diamond with 71.3%.

Brief IA — L'actualité IA en français

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