Google has a new middle child, and it is not shy. On November 18, Google DeepMind released Gemini 4 Neon, a fast reasoning model that sits between the flagship Gemini 4 Carbon and the budget workhorse Gemini 4 Hydrogen. Google says Neon brings “frontier-level intelligence with fast end-to-end task completion” to coding, knowledge work and science, at $3 per million input tokens and $15 per million output tokens. Independent testers at Artificial Analysis put it third on their Intelligence Index, and it was quickly one of the most discussed AI topics on X. Here is what Neon is, what the numbers really show and where the caveats are.
Quick Take
- What: Gemini 4 Neon is Google’s new fast reasoning model with a 1 million-token context window and four thinking levels (low, medium, high and xhigh).
- Price: $3 per million input tokens, $0.15 for cached input and $15 per million output tokens, 25% below Gemini 4 Carbon’s standard price.
- Performance: Artificial Analysis scores Neon (xhigh) at 61 on its Intelligence Index, one point ahead of Carbon and third overall behind GPT-6.2 Astra and Claude Fable 5.2.
- Availability: Gemini API in Google AI Studio, Gemini for business, Google Antigravity for paid Pro and Ultra subscribers, and the Gemini app for Google AI Pro and Ultra subscribers.
- The catch: Most rankings come from Google’s own launch materials or a handful of benchmarks, and early hands-on tests show it is not always the fastest option.
In plain English
Google released a new version of its Gemini AI that is meant to be almost as smart as its best model but quicker and cheaper to run. It is aimed at demanding jobs like writing software, analyzing documents and working through multi-step tasks on your behalf. People who pay for Google’s AI Pro or Ultra plans can pick it in the Gemini app today. For most users, the practical benefit is getting more careful answers without burning through usage limits as fast.
What happened
Google announced Neon across its official channels on the evening of November 18 (European time). Google DeepMind summed up its position in the lineup:
Gemini 4 Neon is rolling out now.
— Google DeepMind (@GoogleDeepMind) November 18, 2026 · View on X
Sitting between Carbon and Hydrogen, it brings frontier intelligence and fast task completion for complex software engineering, knowledge work, and science.
Logan Kilpatrick, who leads Google AI Studio, framed it as a balance model:
Introducing Gemini 4 Neon! Pushing the frontier across domains and punching above its model class in a meaningful way. To me, Neon feels like the right balance of intelligence, speed, and usability. Available now across @GoogleAIStudio, the API, and more!
— Logan Kilpatrick (@OfficialLoganK) November 18, 2026 · View on X
According to Google’s launch post, Neon is rolling out in the Gemini API in Google AI Studio and Android Studio, in Gemini for business (via the API and the model dropdown), in Google Antigravity for paid Google AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. In the app, it appears as 4 Neon (“Deeper analysis”) in the model dropdown in Chat and Tasks. A cyber-focused variant, “Gemini 4 Neon for cyber,” is rolling out to vetted defenders in Google’s Fairwind Program over the coming days.
Background: four Gemini 4 models in seven weeks
Neon is the latest step in an unusually fast release cadence. Artificial Analysis calls it Google’s “fourth model in just seven weeks.” Gemini 4 Carbon first launched for cyber defenders and then opened to paying API developers on November 10, while Gemini 4 Hydrogen arrived as the cost-efficient workhorse for high-volume work. Neon fills the gap in between. In Google’s words: “While Carbon tackles your hardest, long-horizon projects and Hydrogen handles fast everyday execution, Neon balances depth and efficiency.”
The timing is also competitive. OpenAI shipped GPT-6.2 Astra and Sol one day earlier, and Neon is priced right between those two.
How it works
Google has not published architecture details beyond its model card, but the launch post describes how Neon is meant to be used:
- Four thinking levels: Low, medium, high (the default) and xhigh let teams trade intelligence against token spend. Google says xhigh “adds deeper, multi-step reasoning for the hardest engineering, analytical, and scientific workloads.”
- Long context, many inputs: Neon takes text, images, video and speech in a single 1 million-token context window, and Google pitches it for turning spreadsheets, PDFs, charts and recordings into structured deliverables.
- Orchestrator role: In Antigravity, Neon can run on its own or act as “a fast, high-judgment Orchestrator directing parallel Hydrogen subagents.” Google says it cuts down on redundant file reads and handles context compaction reliably in long sessions.
- Steerability: Google highlights that Neon ranks #2 on steerability on the new Agent Arena Index, meaning it adapts to corrections in the middle of a task.
What the numbers say
Here is how Neon’s list price compares with its neighbors (per million tokens):
| Model (per 1M tokens) | Input | Output | Note |
|---|---|---|---|
| Gemini 4 Neon | $3 | $15 | Cached input $0.15 |
| Gemini 4 Carbon (standard) | $4 | $20 | Discounted to $2/$10 until November 30 |
| GPT-6.2 Sol | $2 | $10 | OpenAI, released November 17 |
| GPT-6.2 Astra | $8 | $40 | OpenAI, released November 17 |
Artificial Analysis, which tests models independently, reports that Neon (xhigh) costs $6.44 per Intelligence Index task, about 17% less than Carbon, and takes about 21 minutes per task, about 40% less. The firm notes that Neon actually uses about 15% more output tokens than Carbon, so the savings come purely from lower token prices and a higher output speed of about 110 versus about 57 tokens per second. At the high setting, Neon matches Carbon’s Intelligence Index score of 60 for $3.74 per task, less than half of Carbon’s cost at standard pricing.
Google continues to accelerate its release cadence with Gemini 4 Neon, delivering a small Intelligence gain over Gemini 4 Carbon at a ~17% lower Cost per Task
— Artificial Analysis (@ArtificialAnlys) November 18, 2026 · View on X
On coding agents, Artificial Analysis reports that Neon (xhigh) with the Antigravity CLI scores 59 on its Coding Agent Index, in line with Claude Code running Claude Fable 5.2 (58) and Claude Opus 5.5 (58), at an average cost per task of $6.86 versus $12.61 and $15.47 for the two Claude setups. Google’s own post adds a long list of results, including 70.9% and #2 on the Vals Index, 65.2% on Terminal-Bench 4.1, 46.2% on FrontierCode 2.0 at the default high setting, 71.3% on HLE-Diamond and 39.0% on SWE-sweep inside Antigravity (12.8% on the simpler mini-SWE harness).
On safety, Google says Neon keeps Carbon-level protections against severe misuse and prompt injection, reports a 0.0% attack success rate on Gray Swan’s indirect prompt injection benchmark at its highest thinking level with real-time safeguards, and records shortcuts on 2.45% of tasks in its internal coding cheating evaluation, the lowest of any Gemini 4 model.
Pros and cons
For users
- Pro: Pro and Ultra subscribers get a stronger “Deeper analysis” option, and Google says Tasks on Neon stretch usage limits further than larger models.
- Con: In the Gemini app, Neon is limited to paying Pro and Ultra subscribers, and the growing list of Gemini 4 names (Argon, Carbon, Hydrogen, Neon) makes it harder to know which model to pick.
For developers
- Pro: Near-flagship coding agent scores at a fraction of the cost per task of the leading Claude setups, a 1M context window and fine-grained thinking levels.
- Con: At $3/$15, Neon is pricier per token than Carbon’s temporary $2/$10 rate until November 30, and xhigh runs can be slow in wall-clock time.
For companies
- Pro: Strong results on enterprise-style benchmarks such as legal research, tax work and spreadsheet modeling, plus availability in Gemini for business.
- Con: Many headline numbers are self-reported or chosen by Google, so internal evaluations on real workflows remain essential before switching production traffic.
What X is saying
Reactions were mostly enthusiastic, with plenty of side-by-side tests. Koray Kavukcuoglu, Google DeepMind’s SVP and chief AI architect, wrote that the team “focused on making Neon better to work with: much improved steerability (#2 on Agent Arena Steering), strong multi-step instruction following, and our lowest fabrication rate yet in internal Google @Antigravity testing.”
Hands-on testers were more mixed. Developer @notjazii compared Neon with OpenAI’s cheaper Sol model on the same prompt:
gemini 4 neon vs gpt 6.2 sol
— J A Z I I (@notjazii) November 18, 2026 · View on X
tested both models with same prompt at highest reasoning available and results are kinda same
> neon took 20 mins and cost ~$1.6
> sol took 9 mins and cost ~$0.53
even tho neon feels fast but after testing it against sol, i think sol is way faster
In a separate test against Claude Opus 5.5, the same user reported that Neon took 20 minutes and cost about $3.5 versus 18 minutes and about $7.2 for Opus, and called Neon’s output detail “just crazy good.” Builder @OzanTRAI, who asked Neon to build a small game, wrote: “My take: close to Fable and Opus level. Fable built a similar game for me and still did it better.” Another early tester, @PyrmiGuy, summed it up as “Better than Hydrogen, worse than Carbon. Overall not bad at all.”
Grok, replying to a pricing question on X, summarized: “Gemini 4 Neon API is $3/M input + $15/M output (cached $0.15). On the Coding Agent Index it averages ~$6.86 per task and ranks #1.”
Critical analysis
Neon looks like a genuinely useful product decision rather than a leap in raw capability. Artificial Analysis describes the intelligence gain over Carbon as “small,” and the real story is cost and speed: roughly Carbon-level results for less money and less waiting. For teams running many agent tasks, that matters more than a one-point leaderboard bump.
The caveats are familiar. Google’s launch post cites more than a dozen benchmarks, several of them new or niche, and lists mostly the ones where Neon ranks first or second. Hands-on tests on X show that “fast” is relative: against GPT-6.2 Sol, one tester found Neon slower and about three times more expensive for a similar result. The pricing comparison with Carbon is also time-limited, because Carbon’s discounted $2/$10 rate ends on November 30. And Google’s 0.0% prompt injection result on Gray Swan’s benchmark is a single test; other researchers have shown that differently designed attacks can still get through Gemini models, which is a reminder that safety scores depend heavily on the test.
Finally, the naming is starting to work against users. Four Gemini 4 models in seven weeks is impressive engineering, but customers now have to understand a periodic table of options to choose the right one.
What’s still open
- How Neon performs in independent, longer-running production tests beyond launch-week benchmarks.
- Whether Google will change Carbon’s pricing after the discount ends on November 30, which will reshape the Neon versus Carbon math.
- Which of the leaked or rumored next models (such as Krypton) follow, and where they will sit in the lineup.
- Usage limits for Neon in the Gemini app and Antigravity for Pro versus Ultra subscribers in day-to-day use.
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.
This independent i10x analysis is based on Google’s launch post “Introducing Gemini 4 Neon,” official posts by Google DeepMind, the Gemini app, Google AI Studio and Google leaders on X, and Artificial Analysis’s published evaluation, as of November 19, 2026. Benchmark figures are as reported by Google or Artificial Analysis. Hands-on impressions are individual tests shared on X and are not representative measurements.
Sources (plain text, no links):
- Google AI Studio / Google DeepMind: “Introducing Gemini 4 Neon” launch post (November 18, 2026)
- Google DeepMind: Gemini 4 Neon model card
- Posts on X by @GoogleDeepMind, @GeminiApp, @OfficialLoganK, @koraykv and @sundarpichai (November 18, 2026)
- Artificial Analysis: Gemini 4 Neon evaluation thread on X (@ArtificialAnlys, November 18, 2026)
i10x Perspective
The frontier race is quietly turning into a price-performance race. Neon does not dethrone the biggest models, but it makes near-top intelligence cheap and fast enough to run all day inside agents, and that is where most real AI work is heading. Our advice: test Neon at the high setting first, since that is where Google and Artificial Analysis see the best value, and keep an eye on Carbon’s price after November 30. More on Google’s lineup in our Gemini hub.






