Gemini 4 Neon: Filling the Gap Between Performance and Speed
A2Agent Team ยท 2026-11-19T00:00:00Z
Google released Gemini 4 Neon on November 18, 2026, its fourth Gemini 4 model in about a month. Google calls it a mid-range model, and its model card says it's "adapted from Gemini 4 Carbon." It's meant for work that needs close to Carbon's reasoning but runs too often, or needs answers too quickly, to justify Carbon's cost and speed.
Why Neon exists
Teams building on AI keep running into the same trade-off. Frontier models reason well, but their price and response times are hard to sustain in high-volume production. Fast, cheap models fit the budget, but they make more mistakes on harder work like refactoring across files or deep data analysis.
Neon sits between the two. It has a 1M-token context window and takes text, images, audio, video, charts, spreadsheets and PDFs as input. It offers four thinking levels (Low, Medium, High and Extra High), with High as the default. At Extra High it scores 61 on the Artificial Analysis Intelligence Index, just above Carbon's 60 and third overall.
Where Neon fits in the Gemini 4 lineup
All four Gemini 4 models are named after chemical elements:
| Gemini 4 Hydrogen | Gemini 4 Neon (new) | Gemini 4 Argon | Gemini 4 Carbon | |
|---|---|---|---|---|
| Google's description | Small, fast workhorse for everyday help | Mid-range, "frontier-level reasoning for complex, high-frequency workflows" | "The ideal choice for the majority of software engineering and knowledge work" | Built "for tasks that test the limits of a model's intelligence" |
| Gemini app | "Everyday help" | "Deeper analysis" (Pro and Ultra) | Replaced by Neon | "Code & complex work" (Ultra) |
| Output speed | 167 to 171 tokens/s | About 110 tokens/s (Extra High) | n/a | About 59 tokens/s (High) |
| Price per 1M tokens (input / output) | $1.50 / $7.50 | $3 / $15 | $2 / $10 until Nov 30, then $4 / $20 | $2 / $10 until Nov 30, then $4 / $20 |
| Cached input per 1M | $0.075 | $0.15 | $0.10, then $0.20 | $0.10, then $0.20 |
Two things stand out. Neon replaces Argon in the Gemini app's model picker, so for app users the lineup is now Hydrogen, Neon and Carbon. And Neon's price looks odd for a few days. Until November 30 it costs more than Carbon, which is on an introductory $2 / $10. From December 1, when Carbon and Argon move to $4 / $20, Neon is 25% cheaper than both.
Speed and cost
Measured by Priors, Neon outputs about 110 tokens per second at Extra High. Artificial Analysis lists Carbon at 59 tokens per second on High, so Neon generates output about 1.9 times faster. Hydrogen is faster still, at around 170.
Neon also takes a long time to start at its highest setting: about 57 seconds to the first token at Extra High, per Priors. That's fine for background agent work and slow for anything a person is waiting on.
On cost, Google says Neon is 32% cheaper per task than Carbon at the default High setting. Artificial Analysis found a smaller gap at Extra High, about 17% cheaper than Carbon at High. Glean, an early tester, said Neon matched or beat Carbon in its evaluations "at 3x the speed and 60% of the cost."
Benchmarks
Google compared Neon with Claude Opus 5.5 and GPT-6.2 Sol:
| Benchmark | Gemini 4 Neon | Claude Opus 5.5 | GPT-6.2 Sol |
|---|---|---|---|
| FrontierCode 2.0 | 46.2% | 47.3% | 44.4% |
| Terminal-Bench 4.1 | 65.2% | 65.8% | n/a |
| Terminal-Bench-Science 0.1 | 70.5% | 63.3% | 68.0% |
| AutomationBench (Zapier) | 56.3% | 42.5% | 41.3% |
| RiemannBench | 84.0% | 80.0% | 80.0% |
| HLE-Diamond | 71.3% | 67.9% | 64.0% |
| Chartography | 73.5% | n/a | 75.1% |
On coding, Neon is within about a point of Opus 5.5 on FrontierCode 2.0 and Terminal-Bench 4.1, at 25% lower token prices than Opus. Its clearest leads are on AutomationBench, where it beats both rivals by about 14 points, and on science and math, where it leads on Terminal-Bench-Science, RiemannBench and HLE-Diamond. GPT-6.2 Sol is ahead on chart understanding.
Google also reports 70.9% on the Vals Index, second overall, with first place on legal research and tax agent tasks.
Agent setup changes the numbers a lot. On SWE-sweep, Neon scored 12.8% with the minimal mini-SWE agent and 39.0% inside Google's Antigravity, with the same model in both. If you test Neon in your own agent framework, expect results somewhere in that range rather than at the top of Google's chart.
The Artificial Analysis Intelligence Index puts these models in this order:
| Model | AA Intelligence Index |
|---|---|
| Gemini 4 Neon (Extra High) | 61 |
| Gemini 4 Carbon (High) | 60 |
| Claude Opus 5.5 | 59 |
| GPT-6.2 Sol (max) | 57 |
| Gemini 4 Hydrogen (Extra High) | 54 |
Neon for Cyber
Google is also releasing Neon for Cyber, a version for defenders, to eligible participants in its Fairwind Program. It scores 90.5% on a vulnerability-finding evaluation and 70.7% on a patching test. Enterprises request access through their Google Cloud account team.
Neon with other models
In Antigravity, Neon can split a project across several Hydrogen agents working in parallel, with Neon planning and Hydrogen executing. That's the practical use of the lineup: Carbon for the hardest long projects, Neon for planning and complex day-to-day engineering, and Hydrogen for the many quick tasks underneath.
Availability
Neon is available in the paid Gemini API, Google AI Studio, Android Studio, the Gemini app for AI Pro and Ultra subscribers, and Google Antigravity.
Build multi-model workflows on A2Agent
A2Agent gives you DeepSeek, GLM, Kimi, MiniMax, Qwen and other models behind one API key, with OpenAI-compatible and Anthropic-compatible endpoints, so you can set up the same planner-and-executor split with a stronger model on planning and a cheap, fast one on execution. See the Models page for current availability and prices. New accounts get $5 in free credit.
Sources
- Google launches Gemini 4 Neon with Opus 5.5-level coding performance at 25% lower API token prices (VentureBeat, November 18, 2026)
- Google launches Gemini 4 Neon with Carbon quality at Flash speeds (9to5Google, November 18, 2026)
- Google launches Gemini 4 Neon, pricing it at $3 per million input tokens (CASRAI)