Gemini 4 Neon is the most interesting model Google has shipped this year, and I can’t wait to point it at real work. It’s also the fourth Gemini 4 model in about seven weeks, and the first one, Argon, is already retired. That second fact matters more to how you build than any benchmark on the launch page.
Why Neon is worth the hype
Neon launched November 18 at $3 per million input tokens and $15 per million output, with cached input at $0.15. That’s a 95% discount on anything you send twice. You get a one-million-token context window and four thinking levels, and it takes spreadsheets, PDFs, charts, audio and video in the same context.
The numbers that caught my eye aren’t the top-line scores. They’re the cost-adjusted ones:
- Cheaper than the flagship. Google says Neon costs 32% less per task than Gemini 4 Carbon at its default high setting. Artificial Analysis has it matching Carbon’s Intelligence Index score at high effort.
- Near the top for half the price. On the Vals Index it’s second at 70.9%, about 0.4 points behind the leader, at less than half the cost per task.
- Built for agents. It debuted second on Agent Arena and first on Zapier’s AutomationBench. In Antigravity it can direct a crew of parallel Hydrogen subagents, which is the architecture I’d actually want for long jobs.
Cursor and GitHub Copilot had it on day one. Developers will have it in their hands this week, not next quarter.
The catch is the calendar
Argon came out at the end of September. Carbon followed about four weeks later, Hydrogen two weeks after that, and Neon six days after Hydrogen. By the time Neon arrived, Argon had been retired. That’s roughly a six-week lifespan for a frontier model.
If your app hardcodes a model name, your shelf life is now measured in weeks. I’ve lived through enough enterprise platform migrations to know what that turns into: somebody’s weekend, a pile of prompts tuned for a model that no longer exists, and a cost line that moves every time the default changes.
The teams that win with this release won’t be the ones who switch to Neon fastest. They’ll be the ones who can switch to anything quickly:
- Put a router in front of the model. Pick a model per task, not per app. Neon for the heavy lifting, something like Hydrogen for the fast cheap stuff.
- Keep your own evals. A benchmark tells you how a model does on someone else’s work. Twenty real tasks from your own backlog tell you whether to switch.
- Design for the cache. At $0.15 versus $3, how you structure context is now a bigger cost lever than which model you pick.
I’m going to spend the weekend throwing my own agent workloads at Neon, and I expect to be impressed. I also expect to be doing this again before Christmas. Build for the swap, because the next model is already on the way.
Sources
- Introducing Gemini 4 Neon, Google, Nov 18, 2026
- Gemini 4 Neon Model Card, Google DeepMind, Nov 18, 2026
- [AINews] Gemini 4 Neon: GDM’s Opus/Sol Contender, Latent Space, Nov 19, 2026
- Google DeepMind Unveils Gemini 4 Argon, MarkTechPost, Sep 30, 2026
- Google launches Gemini 4 Hydrogen, Techmeme, Nov 12, 2026
