Google Launches Gemini 4 Neon, Topping a Key Coding-Agent Leaderboard at Lower Cost
News highlights
- Google released Gemini 4 Neon on November 18, 2026, a mid-tier model that sits between Gemini 4 Hydrogen and the flagship Gemini 4 Carbon.
- API pricing is $3 per million input tokens and $15 per million output tokens, with cached input at $0.15, a 1 million-token context window and no free API tier.
- Independent benchmarker Artificial Analysis ranks Neon third on its Intelligence Index and first on its Coding Agent Index, at a lower cost per coding task than Anthropic’s Claude Opus 5.5 and Claude Fable 5.2.
- Neon costs 50% more per token than Carbon until November 30. From December 1, when Carbon’s introductory rate ends, Neon becomes 25% cheaper.
- A cybersecurity version, Gemini 4 Neon for cyber, is going only to vetted defenders through Google’s Fairwind Program.
Google Fills the Middle of Its Gemini 4 Lineup
Google has released Gemini 4 Neon, a new model in its Gemini 4 family that it says brings near-flagship reasoning to coding, analysis and science work at a lower cost per job. The model went live on November 18 on the paid Gemini API. Early independent testing supports much of Google’s pitch, though a pricing quirk in Google’s own lineup means the savings depend on when, and how, a team uses it.
Neon sits in the middle of Google’s range. Hydrogen handles fast, everyday tasks, while Carbon is reserved for the hardest long-running work. Neon has a 1 million-token context window and, according to Artificial Analysis, accepts text, images, video and speech while returning text only. That breadth of multimodal input lets a single request combine, for example, a recorded call, a contract PDF and a spreadsheet. Teams can choose from four thinking levels (low, medium, high and xhigh), trading reasoning depth against speed and spend.
Independent Tests Back the Coding Claims
Artificial Analysis, which runs its own model evaluations, scored Neon 61 on its Intelligence Index at the xhigh setting. That places it third overall, one point above Carbon and behind OpenAI’s GPT-6.2 Astra and Anthropic’s Claude Fable 5.2, both measured at their maximum settings.
The stronger result is in agentic coding. The Artificial Analysis Coding Agent Index tests each model inside its maker’s own terminal agent, across software engineering, technical Q&A and terminal tasks. Running in Google’s Antigravity CLI, Neon placed first at 58.6, just ahead of Claude Fable 5.2 in Claude Code at 58.47, according to a leaderboard reading published by OrcaRouter.
Score barely separates the leaders; cost does. Neon averaged less than half the per-task cost of Claude Fable 5.2 at its xhigh setting, and roughly half that of Claude Opus 5.5. Only GPT-6.2 Astra ran cheaper, at a noticeably lower score. Neon is also quick: Artificial Analysis clocked its output at about 110 tokens per second, against roughly 57 for Carbon.
Cheaper Per Task, Pricier Per Token Until December
Google says Neon costs 32% less per task than Carbon at its default high setting. Artificial Analysis, testing at xhigh, measured a smaller gap of about 17%: $6.44 per task against $7.76 for Carbon. It also found that Neon produces roughly 15% more output tokens than Carbon, so the saving comes from Neon’s lower list price rather than from leaner output.
That comparison uses Carbon’s standard rate. Right now, Carbon is on an introductory price that runs through November 30, which makes Neon the more expensive of the two per token for the rest of the month.
| Price per 1M tokens | Gemini 4 Neon | Gemini 4 Carbon |
|---|---|---|
| Input, through Nov 30 | $3 | $2 |
| Output, through Nov 30 | $15 | $10 |
| Input, from Dec 1 | $3 | $4 |
| Output, from Dec 1 | $15 | $20 |
Teams already running Carbon in production face a budget change either way. Staying on Carbon means token costs double in December; moving to Neon means paying more per token today for a lower rate next month.
Against rivals, VentureBeat reports that Neon’s rates sit 50% above Anthropic’s Claude Sonnet 5.5 and OpenAI’s GPT-6.2 Sol, both at $2/$10, and 25% below Claude Opus 5.5 at $4/$20.
Built for Agents, Sensitive to Tooling
Google positions Neon for long-running, tool-heavy work. In Google Antigravity, its agentic development platform, Neon can act as an orchestrator that hands parallel subtasks to Hydrogen agents, a pattern common in multi-agent architectures. Google says the model also cuts redundant file reads and handles context compaction reliably in long sessions.
Google’s own numbers show how much the surrounding software matters. On the SWE-sweep evaluation, Neon scored 12.8% with the mini-SWE agent harness and 39.0% inside Antigravity, with no change to the model itself.
For knowledge work, Google reports that Neon placed second on the Vals Index at 70.9% and first on Vals’ legal research and tax agent benchmarks. Those placements come from Google’s announcement; VentureBeat noted that Vals’ public leaderboard snapshot predated Neon’s release.
Where Neon Is Available
Developers can use Neon through the paid tier of the Gemini API in Google AI Studio and in Android Studio, Google’s official IDE for Android app projects. There is no free API tier. Enterprises can use it through Gemini for business, via the API or the model drop-down. Paid Google AI Pro and Ultra subscribers can select Neon in the Gemini app and use it in Antigravity.
A separate version, Gemini 4 Neon for cyber, targets finding, validating and patching code vulnerabilities. Google says it will reach eligible members of its Fairwind Program over the next few days, with access requests handled through Google Cloud account teams.
What It Means for Tech Leaders
Neon is Google’s fourth Gemini 4 model in seven weeks, by Artificial Analysis’s count, and it arrives a day after OpenAI’s GPT-6.2 Astra and Sol. For companies bringing agentic AI into enterprise workflows, the useful question is no longer which model leads a benchmark this week, but what a finished workload costs on a given model, in a given agent setup.
Test candidates inside your own agent stack rather than relying on leaderboard ranks, and revisit your AI agent cost forecasts before Carbon’s new rates apply on December 1. Weighing which model belongs where in your stack? Our enterprise AI development team can help.
