Google Launches Gemini 4 Neon, Third at Artificial Analysis

Google DeepMind has launched Gemini 4 Neon, ranked third by Artificial Analysis behind GPT-6.2 Astra and Claude Fable 5.2. The model surpasses Carbon in most evaluations, but its pricing advantage does not apply to the promotional price of the latter.
Visual Results and Cybersecurity Tests
Artificial Analysis notes significant advancements in visual processing. For GDP.pdf, Neon achieves the best result in "high" mode, at 33.4%, five points ahead of Carbon; however, in "xhigh" mode, which is higher, it drops to 28.0%. For presentation quality assessed by AA-Briefcase, it outperforms Opus 5.5 with 1771 Elo points.
Google also claims the top spot in certain categories of the Vals Index, as well as on AutomationBench from Zapier, with 56.3%, and on the scientific test RiemannBench, with 84.0%.
For cybersecurity specialists, Google is launching an accessible version through the Fairwind Program. It scores 90.5% on SEC-bench Pro and 51.26% on WebExploitBench v1.2, at a price half that of competing models. Google further asserts that Neon stops 100% of indirect instruction injections tested on Gray-Swan.
Pricing and Programming Task Costs
On Google AI Studio as well as Gemini for Business, Neon is priced at three dollars per million incoming tokens, or $0.15 with caching, and fifteen dollars per million outgoing tokens. The displayed reduction is 25% compared to Carbon's usual rate. However, Neon is 50% more expensive than Carbon at the promotional rate, which is maintained until the end of November.
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
Artificial Analysis estimates a task in "xhigh" mode at $6.44 for Neon, compared to $7.76 for Carbon without discount: the gap reaches about 17%. Since Neon generates about 15% more output tokens per task, this savings is entirely due to their lower unit price. At the "high" setting, the second highest level, it matches Carbon's intelligence score for a per-task expenditure reduced by half. The benchmark takes about 21 minutes, compared to around 36 minutes for Carbon, thanks to a throughput of about 110 tokens per second versus 57, nearly double.
In programming, the Coding Agents Index assigns Neon set to "xhigh" in Google Antigravity 59 points, compared to 55 for Carbon and 58 each for Claude Fable 5.2 and Claude Opus 5.5. The average cost of a task is $6.86 with Neon, compared to $12.61 for Opus and $15.47 for Fable. On SWE-sweep, Google claims that using Antigravity triples Neon’s success rate, which reaches 39.0%.
Ranking and Position in Google's Range
The table published by Google only considers Claude Opus 5.5 and GPT-6.2 Sol as competitors. It does not include the models from Anthropic and OpenAI that precede Neon in Artificial Analysis. At the maximum "xhigh" setting, the Intelligence Index gives Neon 61 points, compared to 60 for Carbon.
Neon is the fourth model launched by Google DeepMind in seven weeks. Google positions Neon between Carbon and Hydrogen in its offering. Its context capacity reaches one million tokens, supporting text, images, audio, and video.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.