Gemini 4 Neon speeds up use cases without a major intelligence leap
Google made Gemini 4 Neon generally available on November 18, 2026. The model focuses mainly on speed and cost, with clear gains in coding, while the intelligence leap remains modest.
Google DeepMind made Gemini 4 Neon generally available on November 18, 2026, for software development, document work, and science use cases. The model sits between Gemini 4 Carbon and Hydrogen, with a one million token context window for text, image, audio, and video. Its main appeal lies less in a major reasoning leap than in faster, cheaper responses, which quickly matters when requests follow one another.
Une montée en puissance qui se joue surtout sur la vitesse
On the Artificial Analysis benchmark, Gemini 4 Neon ranks third, behind OpenAI and Anthropic. At the highest reasoning level, “xhigh,” it scores 61 points on the Intelligence Index, one point more than Gemini 4 Carbon, but without radically changing the hierarchy. However, Neon runs at about 110 tokens per second, compared to 57 for Carbon, significantly reducing execution time. On the same test suite, the average time drops from about 36 minutes to 21 minutes. Google’s model thus moves faster. And in this market, speed often ends up making the difference between a tool you test and a tool you keep.
Cost follows the same logic. According to Artificial Analysis, one million input tokens cost 3 dollars via Google AI Studio and Gemini for Business, or 0.15 dollar in cache; one million output tokens cost 15 dollars. That’s 25% less than Carbon’s standard rate, but 50% more than its promotional price still in effect until November 30, 2026 [to be verified]. Yet, the calculation does not come from lower token consumption: Neon uses about 15% more tokens than Carbon on some tests. The savings therefore mainly come from the sticker price, not magical efficiency.
Le codage, terrain où Neon prend l’avantage
In programming tasks, Gemini 4 Neon proves particularly competitive. In Antigravity, it reaches 59 points on the Coding Agent Index in xhigh mode, ahead of Carbon at 55 points, and ahead of Claude Fable 5.2 as well as Claude Opus 5.5, both at 58 points. The average cost of a coding task drops to 6.86 dollars, compared to 12.61 dollars for Opus and 15.47 dollars for Fable. Google also highlights an increase in success rate in its Antigravity environment: on SWE-sweep, the rate climbs to 39.0%.
This gain does not tell the whole story. Artificial Analysis notes that higher effort settings do not always yield the best result. On the GDP.pdf test, Neon in high mode reaches 33.4%, ahead of Carbon at 28.4%, but xhigh mode falls back to 28.0%. In other words, pushing the engine harder does not always help. The model remains good, but it does not always obey the reflex “more effort = better result,” which reassures both product teams and spreadsheets.
Vision, agents et sécurité : un modèle plus large que le texte
Beyond code, Neon shows solid results in visual analysis and agent work. On AA-Briefcase, it scores 1771 Elo points in presentation quality, ahead of Claude Sonnet 5.5 at 1726. On the GDP.pdf document test, it rises to 33.4% in high mode. Google also positions it at the top in certain Vals Index categories, on Zapier AutomationBench at 56.3%, and on scientific tests like RiemannBench at 84.0%.
Security is part of the package. Google is preparing a cyber variant via the Fairwind Program, presented as achieving 90.5% on SEC-bench Pro and 51.26% on WebExploitBench v1.2, at half the price of competing models. According to the Model Card, Neon blocks 100% of indirect prompt injection attacks on the Gray-Swan benchmark. The figure is appealing, but it remains linked to a specific test framework [to be verified]; security teams rarely look at a single score before signing.
Key points
- Gemini 4 Neon became generally available on November 18, 2026.
- Google positions it between Carbon and Hydrogen.
- Artificial Analysis ranks it third behind OpenAI and Anthropic.
- The xhigh mode reaches 61 points on the Intelligence Index.
- The average cost of a coding task drops to 6.86 dollars.
- Speed rises to about 110 tokens per second.
By the numbers
- 1 million tokens — context window available for text, image, audio, and video, according to Google.
- 3 dollars — price for one million input tokens via Google AI Studio and Gemini for Business.
- 15 dollars — price for one million output tokens via Google’s paid offerings.
- 39.0% — success rate on SWE-sweep in Antigravity, according to Google.
- 90.5% — cyber variant score on SEC-bench Pro, according to the Fairwind Program.