Gemini 4 Neon: Google aims for the top shelf at $3 per million tokens
On November 18, 2026, Google unveiled Gemini 4 Neon, its high-end model for code, office work and science. The price is public and so are the rankings, though for now Google is the one telling the story.

Google announced Gemini 4 Neon on November 18, 2026, billing it as its frontier-class model for software engineering, office work and scientific use. One caveat before we go any further: everything below comes from Google’s own blog post, the only source available at the time of writing. The leaderboards it cites are run by third parties, but the vendor picked which ones to put in the shop window.
The price, the one line nobody argues with
Neon costs $3 per million input tokens and $15 per million output tokens. Cached input tokens drop to $0.15 per million, a 95% discount. The context window stretches to one million tokens, into which Google promises you can pile spreadsheets, PDFs, charts, audio and video. That is roughly your team’s shared drive, with one difference: somebody would finally read it.
The rollout started the same day: the Gemini API in Google AI Studio, Gemini for enterprise, Google Antigravity for paying AI Pro and Ultra subscribers, and the Gemini app for those same subscribers.
Third place, and proud of it
On general intelligence, Google claims a score of 61 on the Artificial Analysis Intelligence Index, at the xhigh thinking level. That puts it third overall, behind models the post describes as bigger and slower, without naming them. Launch posts rarely stop to polish a bronze medal, so credit where it’s due. The underlying pitch is the trade-off between capability and speed.
The rest of the trophy cabinet, still according to Google:
- second on the Vals Index (70.9%), and first on its legal research and tax agent tests;
- first on the Artificial Analysis Coding Agent Index (58.6), with 65.2% on Terminal-Bench 4.1;
- second on the new Agent Arena Index, with a score of 1336 built on more than 4,200 real agent sessions;
- first on Zapier’s AutomationBench (56.3%) and on Andon Labs’ Vending-Bench 2, a long-running business management simulation, with $16,843;
- top of Terminal-Bench Science (70.5%), RiemannBench (84.0%) and HLE-Diamond (71.3%).
What it changes if you write code
Beyond the scores, Google is selling the model’s behaviour. Neon is said to be built for long agent sessions heavy on tool calls: fewer pointless re-reads of files, a progress update along the way, and context compaction advertised as reliable. It can work solo or act as an orchestrator, running Hydrogen sub-agents in parallel inside Antigravity.
Two more promises will ring a bell if you have ever tried to grab the wheel from an agent mid-task. The model ranks second for steerability on the Agent Arena Index, meaning its ability to take a correction from you while the job is under way. Google also reports its lowest fabrication rate to date, measured in internal testing in Antigravity. Internal, so file it accordingly.
Against the competition, the other voices are missing
The comparison everyone wants is also the thinnest part of the announcement. The post places Neon relative to comparable models and to bigger ones, without giving the names or the scores of the opposition in the passages that carry numbers. No independent feedback is available yet to confirm or temper the picture. The indexes cited exist outside Google, and you will be able to check the rankings there. The selection, though, is still the press release’s.
Until then, the only number you can verify yourself is the one on your invoice. Conveniently, it is also the one benchmark that reads the same no matter who is holding the chart.
Sources (1)
- Gemini 4 Neonblog.google
Written with AI assistance from the sources cited above, then reviewed and approved before publication by Sébastien Soulier.


