Google launches Gemini 4 Neon at $3/$15, a middle tier between Carbon and Hydrogen
Neon has a 1 million-token context window and costs more per token than Carbon until Carbon's introductory price ends on November 30.
Google released Gemini 4 Neon on November 18, 2026, the fourth Gemini 4 model in roughly a month. In its launch post, Google positions Neon between Gemini 4 Carbon, which it reserves for the hardest long-horizon work, and Gemini 4 Hydrogen, its fast everyday model. The pitch is close-to-frontier reasoning for coding, analysis and enterprise workflows at lower latency and lower cost per task than Carbon.
For developers the most concrete facts are the price and the limits. Neon costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million, which Google describes as a 95% discount. The context window is 1 million tokens. Google's launch post does not describe the Neon price as introductory.
Where it sits on price
Neon is not the cheapest way to reach Gemini 4 today, and that is easy to miss. Carbon and Argon are still on an introductory rate of $2 input and $10 output per million tokens. Google's Carbon post says that period ends on November 30, 2026, after which Carbon moves to $4 and $20. Hydrogen is $1.50 and $7.50 across its full 1 million-token window. The Decoder laid out the same comparison in its launch coverage.
- Until November 30: Neon ($3/$15) costs more per token than Carbon ($2/$10).
- From December 1 at the announced rates: Neon is 25% cheaper per token than Carbon ($4/$20).
- Hydrogen ($1.50/$7.50) costs half as much per token as Neon throughout.
Google also argues on cost per task rather than cost per token. It says that at Neon's default high thinking level, Neon pairs a high Artificial Analysis Intelligence Index score with 32% lower cost per task than Carbon, and that it costs roughly 50% to 80% less per task than top competing frontier models on the Vals Index and Agent Arena. Those are Google's own calculations. They depend on how many tokens each model spends on a task, and Google does not say in the post which Carbon price it used.
What Google claims on capability
Most of the launch post's numbers point to third-party leaderboards, which is a step up from purely internal testing, but Google picked which ones to cite. According to Google:
- Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, placing #3 overall.
- It ranks #1 on the Artificial Analysis Coding Agent Index (58.6) and scores 65.2% on Terminal-Bench 4.1.
- It ranks #2 on the Vals Index (70.9%) and #2 on Arena's new Agent Arena Index (1336), including #2 on steerability.
- On SWE-sweep it scores 12.8% in the mini-SWE harness and 39.0% inside Google Antigravity, which shows how much the harness affects the result.
Promptea could not open the Artificial Analysis or Vals leaderboard pages from its research environment, so the figures above are reported as Google states them and have not been checked against the leaderboards. The Decoder noted at publication time that independent benchmark results were still to come.
Context, output and the model card
The model card adds a few details the blog post leaves out. Neon is "adapted from Gemini 4 Carbon" and shares its architecture, training data and safety policies. Its knowledge cutoff is March 2026. It accepts text, images, audio and video in a combined 1 million-token input-output context window. By default it returns up to 64,000 output tokens per synchronous call, and up to 1 million through background long decode. The full Safety & Security Report for Neon is described as available "in the coming days" rather than published at launch.
On safety, Google says Neon keeps Carbon-level protections against cyber and biosecurity misuse while refusing less legitimate work. It reports 1.98% refusals on Terminal-Bench 4.0, 0.0% cheating on Terminal-Bench 4.0 and a 2.45% shortcut rate on its internal coding-cheating evaluation. These are internal evaluations and cannot be checked from outside until the full report is out. A separate Gemini 4 Neon for cyber goes to vetted defenders in Google's Fairwind Program over the next few days.
Availability
Neon is rolling out in the Gemini API paid tier (Google AI Studio), in Gemini for business through the API and model picker, in Google Antigravity for paid Google AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. Google says Tasks in the Gemini app can also run on Neon.
What changes for model selection
For teams already on Gemini, Neon adds a middle tier. The practical question is not whether it is "better" but which work justifies its price. A reasonable split, based on Google's own positioning, is Hydrogen for high-volume execution, Neon for multi-step coding and analysis where latency matters, and Carbon for the longest and hardest jobs. For the next 12 days Carbon is cheaper per token than Neon, so a migration made purely to save money should wait until the December pricing applies, or be measured on cost per completed task in your own workload rather than on the rate card.
Google also says the four thinking levels scale more evenly than before, so teams can trade intelligence for token spend. In practice that makes the thinking level a cost setting as much as a quality one, and it is worth testing alongside the choice of model.
Why this matters
- It adds a third Gemini 4 price point, so choosing a model is now a decision about cost per task, not just quality.
- The pricing order flips on December 1, when Carbon's introductory rate ends, which changes the math for any migration.
- Google's leaderboard claims cite third parties but have not yet been independently confirmed, so they should be tested on your own workloads.
Key takeaways
- Gemini 4 Neon costs $3/$15 per million tokens, $0.15 for cached input, with a 1 million-token context window.
- Carbon is $2/$10 until November 30, then $4/$20; Hydrogen is $1.50/$7.50.
- The model card says Neon is adapted from Carbon, has a March 2026 knowledge cutoff, and returns up to 64K output tokens per synchronous call by default.
- The full Safety & Security Report has not been published yet; a cyber variant goes to Fairwind Program members in the coming days.
Sources
- GooglePrimaryIntroducing Gemini 4 Neonblog.google
- Google DeepMindPrimaryGemini 4 Neon — Model Cardstorage.googleapis.com
- Google DeepMindPrimaryGemini 4 Carbon is now availabledeepmind.google
- Google DeepMindPrimaryIntroducing Gemini 4 Hydrogendeepmind.google
- The DecoderGoogle ships Gemini 4 Neon, its fourth new Gemini model in a monththe-decoder.com
- model selection
- api pricing
- context windows
- coding agents
- prompt caching
- gemini
- Google DeepMind
- Gemini 4 Neon
- Gemini 4 Carbon
- Gemini 4 Hydrogen
- Gemini 4 Argon