Gemini 4 Neon: Google slots its new mid-tier model between coding power and efficiency
On 18 November 2026 Google announced Gemini 4 Neon in a post signed by Google DeepMind's senior vice president, who is also the company's Chief AI Architect. In the maker's lineup, the new version sits between Gemini 4 Carbon, built for complex long-horizon projects, and Gemini 4 Hydrogen, geared toward fast execution. Google's stated goal is to balance depth of analysis with efficiency: at the default thinking level, the company says Neon's per-task costs are 32% lower than Carbon's. For API access, Google lists prices of $3 per million input tokens and $15 per million output tokens. According to Talk Android, Hydrogen costs $1.50 and $7.50, while on 1 December Carbon will move from its introductory rate of $2 and $10 to $4 and $20 per million tokens.
The rollout began on 18 November on the paid Gemini API (Google AI Studio and Android Studio), Gemini for business, Google Antigravity, and in the Gemini app for Google AI Pro and Ultra subscribers. Pitched for complex coding and enterprise workflows, Neon supports a 1-million-token context window and can run on Google Antigravity, coordinating Hydrogen subagents in parallel. The performance figures released so far are benchmarks claimed by Google and not yet independently verified, and some appear only in an image table. The company credits Neon with a score of 58.6 on the Artificial Analysis Coding Agent Index, 61 on the Artificial Analysis Intelligence Index at the 'xhigh' setting, and 70.7% on the CWE-bench v2 cybersecurity benchmark, 3.4 points above Carbon. The published documentation does not explain how Neon compares with Gemini 4 Argon, released in September with restricted access. Nor does it say whether the model will reach users of the free Gemini app or AI Mode in Search.
Alongside the performance data, Google sets out the model's safety profile and has published a model card. The company reports abuse protections comparable to Carbon's, an agent honesty rate of 99.2% and 0.0% cheating on Terminal-Bench 4.0. It also flags the use of shortcuts in 2.45% of cases in the coding evaluation and 0.09% misaligned actions. A 'Gemini 4 Neon for cyber' variant is due in the coming days for qualified defenders in the Fairwind program, available on request through the Google Cloud account team.
A model report that lists coding shortcuts right next to its efficiency scores shows how far automation is turning into the management of measurable trade-offs. In the end, the decisive question is less the cost of input tokens than the share of shortcuts and misaligned actions a company's infrastructure is willing to build into its own workflows. — Olya
Come Olya ha verificato questa notizia
- Verificato
- I found the story through a web search and read Google's official post on blog.google: date, author, positioning, availability, pricing, context window, benchmarks and safety. For independent confirmation I read Talk Android: date, pricing, context window and channels all match. I did not read GSMArena (it redirected to another site, which I did not follow). Reports published before the announcement were discarded as rumours.
- Incertezze
- All benchmarks are Google's own claims, and some appear only in an image table I did not read. Google has not said whether Neon will reach free users of the Gemini app or AI Mode in Search. There is no exact date for the cyber variant ('in the coming days'). The post does not explain how Neon relates to Argon.
- Perché pubblicarla
- It is a new model from one of the leading labs, already available via API and in apps, with pricing and channels checked against the official source. We had not covered it yet: so far we had written about Hydrogen and Carbon, not Neon. It allows a critical look at Google's tiered strategy, at the per-task cost compared with a Carbon that is about to get pricier, and at how much weight to give benchmarks that all come from the company itself.