Gemini 4 Neon launches at $3/$15 per million tokens with a 1M context window, slotting between Hydrogen and Carbon. Here's what the pricing and benchmarks actually mean.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
Google shipped its fourth new Gemini model in six weeks on November 18, 2026: Gemini 4 Neon, priced at $3 per million input tokens and $15 per million output tokens, with the same 1 million token context window as the rest of the Gemini 4 family. It slots in between Gemini 4 Hydrogen (cheap and fast) and Gemini 4 Carbon (the reasoning heavyweight), and Google's pitch is blunt: Carbon-level quality at something closer to Flash speed, for less per task.
That's the official line. After reading through Google's launch post, the 9to5Google writeup, and the Hacker News thread arguing about it, here's what actually holds up and what needs a correction before you believe it.
Neon is the third model Google has shipped under the Gemini 4 name this quarter, following Hydrogen (Nov 12) and sitting opposite Carbon (Oct 28) and Argon (Sept 30) in the lineup. Google says it's adapted from Carbon rather than trained from scratch, which lines up with how it performs: strong on the same coding and reasoning benchmarks Carbon leads, just faster and noticeably cheaper to run per task.
Google's own framing of the three-model family, straight from the Gemini blog: Carbon handles the "hardest, long-horizon projects," Hydrogen covers "fast everyday execution," and Neon "balances depth and efficiency" for complex work that doesn't need Carbon's full weight. In the Gemini app, the model picker now reads Auto, Everyday help (Hydrogen), Deeper analysis (Neon), and Code & complex work (Carbon). Argon is gone from that list, and so is Gemini 3.1 Pro, which Google finally retired on this launch.
A quick spec sheet:
| Model | Input (per 1M) | Output (per 1M) | Cached input |
|---|---|---|---|
| Gemini 4 Hydrogen | $1.50 | $7.50 | $0.075 |
| Gemini 4 Neon | $3.00 | $15.00 | $0.15 |
| Gemini 4 Carbon (through Nov 30) | $2.00 | $10.00 | $0.10 |
| Gemini 4 Carbon (from Dec 1) | $4.00 | $20.00 | $0.10 |
Here's the part the launch post doesn't lead with: right now, today, Neon is more expensive than Carbon. Carbon's introductory pricing runs through November 30, and at $2/$10 it's cheaper per token than Neon's $3/$15. The "25% cheaper than Carbon" framing you're seeing in coverage only becomes true on December 1, when Carbon's price jumps to $4/$20 and Neon's $3/$15 suddenly looks like the discount option.
So if you're choosing today, Neon isn't the budget pick against Carbon. It's the budget pick against what Carbon is about to cost. That's a real distinction, and one I'd want cleared up before anyone tells a finance team Neon is saving money starting this week.
| Benchmark | Score | Rank |
|---|---|---|
| Artificial Analysis Intelligence Index (xhigh) | 61 | #3 overall |
| Artificial Analysis Coding Agent Index | 58.6% | #1 |
| Vals Index v2.2 | 70.9% | #2 |
| Agent Arena Index | 1336 | #2 |
| SEC-bench Pro | 90.5% | — |
| AutoBaxBench (hard tier) | 76% | — |
A 61 is genuinely higher than Carbon's own 59-60 on the same index a few weeks back, but Neon still lands at #3 overall, not #1, because Claude Opus 5.5 and GPT-6.2 Sol have both climbed since Carbon's launch. That's the pace of this market right now. A model can post a better raw score than the thing that topped the chart a month ago and still not be the best model available the week it ships.
Where Neon does lead outright is coding agents and automation-style tasks: #1 on the Artificial Analysis Coding Agent Index, #1 on Zapier's AutomationBench, #1 on legal and tax research benchmarks. Google also claims the lowest fabrication rate of any Gemini 4 model, which matters more than it sounds given Carbon's hallucination rate on factual tests jumped to around 11%.
Yes, and this is the one claim that survives contact with independent numbers instead of needing a correction. Community throughput data from OpenRouter puts Neon at roughly 127 tokens per second against Carbon's 46, close to a 3x gap. Artificial Analysis' own measurement shows a smaller but still clear split, 110 versus 57. Different trackers, different exact figures, same direction: Neon is meaningfully faster, full stop.
Google frames the cost side of that as a 32% lower cost per task than Carbon at the default High thinking setting. I'd treat that number the way I treat most "cost per task" marketing figures: directionally real, exact percentage dependent on your workload. But the speed gap alone is reason enough to try Neon on anything that was timing out your patience on Carbon.
| Hydrogen | Neon | Carbon | |
|---|---|---|---|
| Role | Everyday execution | Deeper analysis | Hardest, long-horizon work |
| AA Intelligence Index | 51-54 | 61 | 59-60 |
| Output speed | Fastest | ~2-3x Carbon | Slowest |
| Price (input/output) | $1.50 / $7.50 | $3 / $15 | $2-4 / $10-20 |
| Free tier access | Yes | No | No |
| Fairwind-gated variant | No | Cybersecurity only | Full model (initially) |
This is a genuinely different shape of lineup than "small, medium, large" priced in order. Neon costs more than Hydrogen, and for the next 12 days, more than Carbon too. What you're actually buying with Neon is Carbon's intelligence tier at something closer to Hydrogen's latency. That's a real and useful thing if your bottleneck has been waiting on Carbon to finish thinking. It is not a cheaper way to do what Carbon already does, at least not yet.
The Hacker News thread covering launch day split into the same three arguments that show up every time Google ships one of these.
First, the naming scheme is getting mocked again, worse this time because there are now four non-sequential element names (Hydrogen, Argon, Carbon, Neon) with no obvious size or capability order. Someone made the point I'd have made myself: ask a random person to rank hydrogen, argon, carbon, and neon by atomic mass, and most can't, and that's before you add "but Carbon isn't actually the biggest model here."
Second, there's real skepticism that the benchmark gains translate into day-to-day coding. One commenter put it plainly: "it's not clear Neon is considerably better than Sonnet/Terra" once you move past the leaderboard and into an actual repo. Someone else raised a sharper question, whether Anthropic's apparent edge in daily use comes from Opus reasoning better or from better product character and interface polish creating the impression of intelligence. That's worth sitting with before you re-point an agent stack over a benchmark chart.
Third, the people who've actually run it seem cautiously positive, with the caveat that it needs to avoid Carbon's habit of running long and verbose if it wants to hold "best coding model" past launch week.
If you're on Carbon today purely because of latency, waiting on long agentic runs, test Neon this week. The speed numbers are the most credible part of this launch, and they solve a complaint people have genuinely had about Carbon.
If you picked Carbon for the Fairwind cybersecurity access, nothing changes. Neon's security variant is separately gated and doesn't replace what Carbon does there.
If you're price-sensitive, don't move yet. Hydrogen is still the cheapest model in the family by a wide margin, and Neon doesn't actually beat Carbon on price until December 1. Mark your calendar, not your model picker.
And if you're already happy with Claude Opus 5.5 or GPT-6.2 Sol in your coding workflow, I wouldn't switch off a benchmark chart alone. Neon's #1 finishes are real, but they're concentrated in coding-agent and automation benchmarks, not the general intelligence index, where it still sits behind both of those models. Google shipping four Gemini 4 variants in six weeks tells you more about their release cadence than it tells you which one wins your specific job.
Gemini 4 Neon is Google's third Gemini 4 model released in six weeks, announced November 18, 2026. It sits between Gemini 4 Hydrogen and Gemini 4 Carbon, offering Carbon-level benchmark scores at roughly 2-3x Carbon's output speed.
$3 per million input tokens and $15 per million output tokens, with a 95% discount on cached input ($0.15 per million). That's more expensive than Carbon's current $2/$10 introductory price, but cheaper than the $4/$20 Carbon will cost starting December 1, 2026.
It depends on the task. Neon scores higher on the Artificial Analysis Intelligence Index (61 vs Carbon's 59-60) and leads on coding-agent and automation benchmarks, but Carbon still handles the longest-horizon, most complex reasoning work Google positions it for.
Hydrogen is Google's cheapest and fastest Gemini 4 model, built for everyday tasks and free to use via auto model selection. Neon costs twice as much, scores higher on reasoning benchmarks, and is paid-only, positioned for "deeper analysis" work Hydrogen isn't built for.
OpenRouter throughput data shows Neon running at roughly 127 tokens per second versus Carbon's 46, close to a 3x gap. Artificial Analysis measured a smaller but similar split, 110 versus 57 tokens per second.
Google AI Pro and Ultra subscribers in the Gemini app and Google Antigravity, plus developers through the Gemini API, Google AI Studio, and Gemini for Business. There's no free tier access, unlike Hydrogen.
If you're tracking the rest of this rollout, see our breakdowns of Gemini 4 Carbon, Gemini 4 Hydrogen and Antigravity's Orchestrator, and Gemini 4 Argon vs Claude Opus 5.5. For how Neon's competitors stack up, check Claude Opus 5.5 review and pricing and GPT-6.2 Astra Sol vs Opus.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Claude Code, Codex CLI, Pi, and OpenCode each win on different models. Runta's eval found harness choice swinging cost up to 4.4x on the same model.
lot-map is a free Claude Code plugin that draws a live map of tool calls, subagents, hooks and context usage beside your chat. Here's how to install and use it.
MAI Code 2 claims 96% of Claude Opus 5.5's coding performance at 5% of the price in GitHub Copilot. We checked the benchmarks, the promo pricing, and real-world reports.
Comments