Google DeepMind released Gemini 4 Neon on November 18, a reasoning model that outscores the Carbon model it launched only eight days earlier — at a lower price.
Neon sits between Hydrogen and Carbon in the Gemini 4 family and tops the Artificial Analysis Intelligence Index at 61, against Carbon’s 60.
Eight days. That is all the time Google DeepMind gave its own Carbon model at the top of its lineup before Neon arrived on November 18 and beat it.
Carbon launched on November 10, Hydrogen followed last week, and Neon landed today, completing a three-model family in under a month. The Register notes that Argon, the family’s first member, came out in late September.
Neon Slots in Between Hydrogen and Carbon
Koray Kavukcuoglu, SVP at Google DeepMind and chief AI architect at Google, described the model in a launch blog post.
Gemini 4 Neon: frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows. Begins rolling out today in the Gemini API in @GoogleAIStudio, Gemini for business, for paid Google AI Pro and Ultra… pic.twitter.com/FMiFCFvejw
In his blog post’s section titled “Where Neon sits in the Gemini 4 family,” Kavukcuoglu explained that Carbon handles the hardest long-horizon projects while Hydrogen manages fast everyday execution, and that Neon “delivers frontier-level reasoning for complex, high-frequency workflows faster and more cost efficiently.”
“Engineered for long-running agent workflows, Neon sustains performance through complex, tool-heavy tasks end-to-end — whether running standalone or acting as a fast, high-judgment Orchestrator directing parallel Hydrogen subagents in Google Antigravity,” Kavukcuoglu said.

Neon lists at $3 per million input tokens and $15 per million output tokens — less than Carbon’s undiscounted rate of $4/$20, which takes effect after November 30. Until then, Argon and Carbon sell at $2/$10, and Neon stays 25 percent cheaper once the promotion ends.
Neon Tops the Index at 61
The Intelligence Index rankings put the new model at the front of Google’s own family, as The Register reports:
| Model | AA Intelligence Index | Price per 1M tokens (in/out) |
|---|---|---|
| Gemini 4 Neon | 61 | $3 / $15 |
| Gemini 4 Carbon | 60 | $4 / $20 |
| Gemini 4 Hydrogen | 54 | — |
| Gemini 4 Argon | 50 | $4 / $20 |
On coding, the two outlets covering the launch disagree on the exact figure: MarkTechPost puts Neon at 58.6 on the AA Coding Agent Index, while Aimagazine rounds it to 59, three points ahead of Carbon, which was released only November 10. Both agree Neon leads that index. Aimagazine adds that on the Vals Index — a single measure of AI’s potential economic impact — Neon ranks second, just behind Anthropic’s Claude Fable 5.2.
Not every benchmark goes Neon’s way. MarkTechPost reports that Claude Opus 5.5 still leads on Terminal-bench 4.1 (65.8 percent versus 65.2 percent) and FrontierCode 2.0 (47.3 percent versus 46.2 percent), and that GPT-6.2 Sol wins Chartography.
Starting today, Gemini 4 Neon is available to Google AI Pro and Ultra subscribers on mobile, desktop, and web. This new model excels at work that requires frontier-level reasoning and fast responses. 🧵
All of Google’s numbers, MarkTechPost said, are self-reported and not independently reproduced.
Neon’s clearest win is long context: 95.2 percent on GraphWalks at 256K to 1M tokens, against 80.1 percent for GPT-6.2 Sol and 66.8 percent for Opus 5.5. It also debuts at number two globally, with a score of 1336 on Arena’s new Agent Arena Index across more than 4,200 real-world agent sessions, and takes first on Zapier’s AutomationBench at 56.3 percent, per 9to5Google.
Google Reports Strong Safety Scores
Google reports that Neon runs on the Gemini 4 safety stack with a 0.09 percent misaligned action rate on its agentic safety suite, passes 99.2 percent of the agent honesty evaluation, and shows a 0.0 percent attack success rate on Gray Swan’s prompt injection benchmark at the highest thinking level. Its 2.45 percent shortcut rate on a coding cheating test is the lowest of any Gemini 4 model, according to MarkTechPost.
Where You Can Use Neon Today
Neon is available now through the Gemini API paid tier in Google AI Studio and Android Studio, in Gemini for Business, in Google Antigravity for paid AI Pro and Ultra subscribers, and in the Gemini app itself.
GitHub announced the same day that the model is generally available in Copilot for Pro+, Max, Business, and Enterprise users, billed at provider list pricing. In the Gemini app, it replaces Argon in the model picker, where 3.1 Pro is now retired and 3.5 Flash-Lite becomes the “Auto” option for paid users.

In the app, Google pitches Neon for tougher everyday jobs — iPhone in Canada describes scenarios like feeding it a utility bill and older statements to explain a spiked charge, or cross-checking receipts to negotiate a refund.
The model reads PDFs, photos, spreadsheets, audio and video together, and can pull details from Gmail and Docs. Neon also powers Tasks in Gemini, a feature now free in Canada, though running it on Neon still requires a Pro or Ultra subscription.
Meet Gemini 4 Hydrogen: built for rapid coding, multimodal synthesis, and everyday workflows. It finishes tasks 2.5x faster than 3.8 Flash and surpasses Argon on the @ArtificialAnlys Intelligence Index. pic.twitter.com/pY6eZlggqD
Cheaper but Better: The Naming Puzzle
The Register observes that Google's element-based names don't show a hierarchy the way OpenAI's celestial scheme does, and that Neon's price — cheaper than Carbon despite the higher index score — makes the lineup harder to read, not easier.
Google’s Antigravity Orchestrator adds another layer: it can start a task on Carbon, write a plan, then hand the work to Hydrogen to save tokens, deciding how a developer’s money gets spent unless the developer pins a model manually.
The release continues a run of model launches from Google DeepMind, the research laboratory founded in London in November 2010 by Demis Hassabis, Shane Legg and Mustafa Suleyman.
Google acquired the company in 2014, and in April 2023 it merged with Google’s Brain division to form Google DeepMind, a move prompted by the public release of ChatGPT. The lab first made headlines in 2016, when its AlphaGo program beat the Go world champion Lee Sedol in a five-game match.
Its work since has ranged from game-playing programs such as AlphaZero to AlphaFold, which set records on protein folding benchmarks and led to the release of more than 200 million predicted protein structures in 2022. Google DeepMind is now responsible for the Gemini family and the open-weight Gemma models.
Whether Neon truly replaces Carbon for the hardest work, or merely blurs the line Google drew between its models, is the question the company’s own framing leaves open. Carbon remains, in Google’s words, “Our most capable reasoning model.”
Earlier coverage: Why AI Coding Agents Fail Self-Verification and How to Fix It