Google Launches Gemini 4 Neon, Slotting a Cheaper Coding Model Between Carbon and Hydrogen

Google released Gemini 4 Neon on November 18, slotting a new model into the Gemini 4 family between its most powerful reasoning model, Carbon, and its fastest, Hydrogen. Neon is pitched as the balance point: frontier-level intelligence at meaningfully lower cost, aimed squarely at software engineering, agentic coding, and knowledge-work tasks that don't need Carbon's full reasoning depth but still demand strong accuracy.
The launch matters less for topping any single leaderboard — Neon ranks third overall on the Artificial Analysis Intelligence Index with a score of 61 at its highest "xhigh" thinking setting, trailing larger, slower models — and more for where it lands on cost. Google says Neon costs 32% less per task than Carbon at its default "high" thinking level while remaining competitive on the benchmarks that matter most for coding and agentic work.
Pricing and Where It's Available
Neon is priced at $3 per million input tokens and $15 per million output tokens, with a steep 95% discount on cached input tokens, bringing that rate down to $0.15 per million. It ships with a one-million-token context window. As of launch, it's available in the Gemini API through Google AI Studio and Gemini for Business, in the Gemini app and Google Antigravity for paid AI Pro and Ultra subscribers, and in Android Studio for developers. A dedicated "Cyber" variant is rolling out separately to trusted defenders under Google's Fairwind Program, continuing the company's pattern of giving cybersecurity researchers early access to specialized model variants.
Where Neon Actually Wins
Google's own benchmark disclosures show Neon topping several agentic and coding-specific indexes outright rather than just trailing close behind larger models. It ranks first overall on Artificial Analysis's Coding Agent Index (58.6) and first on Zapier's AutomationBench (56.3%). In security-specific testing, it hit 90.5% on SEC-bench Pro and 51.26% on WebExploitBench v1.2 — both described by Google as state of the art — while using roughly half the tokens Carbon needs on comparable exploit-analysis tasks, a direct translation of the cost advantage into practical throughput.
Google also highlighted safety metrics alongside capability ones, a pairing the company has leaned into since safety researchers flagged rising "reward hacking" behavior in frontier coding models earlier this year. Neon shows a 2.45% rate of taking shortcuts on coding evaluations — the lowest of any Gemini 4 model to date — a 0.09% misaligned-action rate on Google's agentic safety suite, and a 0% attack success rate on the Gray Swan prompt-injection benchmark at its highest thinking setting.
Why Google Needed a Model in This Slot
The three-tier Carbon/Neon/Hydrogen split answers a problem that's become increasingly common across frontier labs this year: most production AI workloads don't need the most powerful available model, but running everything on the fastest, cheapest tier risks quality regressions on harder tasks. Google's own documentation suggests pairing Neon as a planning layer with Hydrogen as an execution layer inside its Antigravity orchestrator — effectively productizing the model-routing strategy that enterprises have been building themselves to control AI infrastructure costs.
That positioning puts Neon in direct competition with the mid-tier coding and agentic models OpenAI, Anthropic, and Microsoft have all been racing to ship this quarter, each framing its offering around the same pitch: near-frontier capability without near-frontier pricing. Google releasing Neon just three weeks after Carbon, and days before the close of a month that also saw Anthropic's IPO and Microsoft's seven-model MAI lineup at Ignite, underscores how compressed the release cadence among frontier AI labs has become — cost-tiered model families are no longer an occasional strategic move but a standing requirement to stay competitive on price-performance across every segment of enterprise demand simultaneously.
Originally reported by Google. Read the original article for additional details.
View original source