Neon, hands-on12 minGoogle's 'middle' model, Gemini 4 Neon, went through ten hard hands-on tests: a two-piano duet, video effects reverse engineering, an animated biology lesson, a full song, an Unreal Engine sword fight, an Airbnb rebuilt in Blender, brain scans and more. It was faster than its rivals almost everywhere. It was not better everywhere.Read now →

Google's Gemini 4 Neon Broke the Usual Rule: Faster and Cheaper Than Carbon, and on Many Tests Better
Cheaper, faster models are supposed to be worse. Google's new Gemini 4 Neon, released November 18, is cheaper and faster than Carbon and still tops leaderboards from tax law to running a vending business. Here's what it is, what the numbers really say, and the bot-fighting arena where it's being put to the test.
There's an unwritten rule in AI that everyone learns fast: pick two. Fast, cheap or smart. If a model is faster and cheaper, you assume it's dumber.
On Wednesday, November 18, 2026, Google shipped a model that seems to ignore the rule.
It's called Gemini 4 Neon. It's faster than Google's flagship, Gemini 4 Carbon. It costs less per task. And on a long list of independent tests, from legal research to running a simulated business, it scores higher.
That makes Neon one of the most interesting releases of a very crowded month. It also raises a fair question: if the "middle" model is this good, what exactly are you paying extra for?
Here's the full picture: what Neon is, how Google's confusing names actually work, what the benchmarks do and don't show, and the strange bot-fighting arena where early testers are already pitting it against GPT-6.2 Astra and Claude Fable 5.2.
The periodic table of Gemini
Start with the names, because almost everyone is confused by them.
Google named its Gemini 4 models after chemical elements. That's created a running joke that Google now has a "periodic table" of models, and the releases are coming so fast that it may run through a good chunk of the table before the year is out.
On launch day, Google's Logan Kilpatrick spent hours answering what was essentially one question asked over and over: which model is which?
Here's the lineup after Neon, in plain English:
- Gemini 4 Carbon is the heavyweight. Google built it for the hardest, longest projects, the tasks that test the limits of a model's intelligence.
- Gemini 4 Neon is the new middle. It's adapted from Carbon, according to its model card, and tuned to balance depth and speed for complex work you do many times a day.
- Gemini 4 Hydrogen is the speed tier, the everyday replacement for the old Flash models. Still very capable, just optimized for quick execution.
- Gemini 4 Argon, the earlier code-and-analysis model, is being replaced by Neon in the Gemini app, according to 9to5Google.
People keep trying to map these onto Anthropic's lineup. Is Neon the Haiku? Is Hydrogen the Sonnet? The cleaner way to think about it is by job: Carbon for your hardest problems, Neon for serious daily work, Hydrogen for fast routine tasks.
In the Gemini app, the model picker now reads: Auto, Everyday help (Hydrogen), Deeper analysis (Neon), and Code and complex work (Carbon). Gemini 3.1 Pro is finally retired.
Why "better, cheaper, faster" is the whole point
Google's strategy with Gemini 4 seems to be less about raw peak intelligence and more about pushing the Pareto frontier: the curve that shows the best quality you can get at each price.
Normally, moving down that curve means giving up quality. Neon's pitch is that it shoves the whole curve forward instead.
The numbers Google published:
- Price: $3 per million input tokens, $0.15 per million cached input tokens (a 95% discount) and $15 per million output tokens.
- Context window: 1 million tokens, with up to 64,000 output tokens by default and up to 1 million through a background "long decode" mode.
- Cost per task: at its default high thinking level, Google says Neon costs 32% less per task than Carbon.
- Against rivals: roughly 50% to 80% lower cost per task than top competing frontier models on the Vals and Agent Arena leaderboards, and at least 36% cheaper at its maximum setting versus competitors at theirs.
There's a wrinkle in the headline price. Carbon and Argon are on introductory pricing of $2 input and $10 output until November 30, 2026. After that they rise to $4 and $20. So on sticker price today, Neon actually looks more expensive than Carbon. From December, Neon is 25% cheaper per token, and it uses fewer tokens per task.
VentureBeat noted another comparison: Neon's per-token rates are 50% higher than Anthropic's mid-range Claude Sonnet 5.5 and OpenAI's GPT-6.2 Sol, but 25% below Claude Opus 5.5's standard $4 and $20, despite landing within about a point of Opus on two engineering benchmarks.
The scoreboard
Here's where Neon gets genuinely impressive. These are the results Google reported, many from third-party leaderboards:
Overall intelligence
- #3 overall on the Artificial Analysis Intelligence Index, with a score of 61 at its highest ("xhigh") thinking level, just behind larger, slower models.
- #2 on the Vals Index at 70.9%, on the Pareto frontier.
- #2 on Arena's new Agent Arena Index (score 1336) across more than 4,200 real-world agent sessions, and #2 on steerability, meaning how cleanly it adapts when you correct it mid-task.
Professional work
- #1 on Legal Research Bench and Tax Agent Bench.
- #1 on most Surge AI benchmarks across finance, accounting, spreadsheet modeling and complex constraint-following.
- 73.5% on Chartography, a test of reading dense professional charts.
- 75.8% on Agents' Last Exam v1.1, on par with leading models.
Agents and automation
- #1 on Zapier's AutomationBench (56.3%).
- #1 on Andon Labs' Vending-Bench 2, a long simulation of running a vending business, finishing with $16,843.
Science and math
- Best in class on Terminal-Bench Science (70.5%), RiemannBench (84.0%) and HLE-Diamond (71.3%).
Early independent testing points the same way. One tester's private benchmark, run within hours of launch, put Gemini 4 models in three of the top six spots: Claude Fable 5.2 in a league of its own, then GPT-6.2 Astra, then Carbon, then Neon, then Argon, then Claude Opus 5.5. Those results were explicitly very early.
The coding asterisk
There's one area where you should keep your skepticism switched on: coding.
On most coding benchmarks, Anthropic's Claude Fable 5.2 and Opus 5.5 have been close to unbeatable this autumn. Google is now pointing to results where Neon leads:
- #1 on the Artificial Analysis Coding Agent Index (58.6).
- 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default setting.
- On the SWE-sweep test, 12.8% on the basic mini-SWE harness, more than tripling to 39.0% inside Google's own Antigravity coding environment.
That last number is worth sitting with. The same model's score tripled depending on the software wrapped around it. Harness matters enormously, and benchmark claims from a model's own maker deserve independent confirmation, especially in the one category where Anthropic has been so dominant.
Early testers hadn't yet given Neon deep coding tasks at launch. If the coding claims hold, it'd be a big deal. If they don't, Neon is still excellent at almost everything else.
What Google showed it building
Google's launch demos give a feel for what "frontier-level at speed" means in practice. In Antigravity, Neon reportedly:
- Drove Blender through an MCP connection, using the /swarm feature to build a photorealistic 3D lab bench illustrating the Gemini element names, with Bunsen flames for Carbon and Hydrogen and gas discharge tubes for Argon and Neon.
- Wrote a first-person freeride ski simulator from scratch on real Mont Blanc and Chamonix elevation data: 10,200 lines of custom shaders, 120 Hz ski physics, carved tracks, avalanches and spatial audio, with no pre-made assets.
- Recreated Alan Turing's 1936 study at King's College, Cambridge, in interactive 3D from archival photos, with a working Universal Turing Machine on the desk.
- Turned a single image into a 100,000-cell Google Sheet that uses more than 500,000 live formulas to do edge detection, map depth and relight the scene, with no embedded images or scripts.
Antigravity's Orchestrator can now also pair Neon as the planner with Hydrogen as the executor in a single session, so you get Neon's judgment directing many fast, cheap workers.
Neon for cyber defense
Neon is also pitched as a security model, and here Google's claims are specific:
- AutoBaxBench: more secure code than Carbon at every difficulty, state of the art on the hard tier (76%).
- WebExploitBench v1.2: state of the art at 51.26%, at half the cost of large competing models.
- SEC-bench Pro: 90.5% at finding vulnerabilities, at roughly half the cost per challenge.
- CWE-bench v2: 70.7% at patching vulnerabilities, a 3.4% uplift over Carbon.
The specialist "Neon for cyber" version is rolling out only to vetted defenders through Google's Fairwind Program.
The timing is telling. On the same morning, Anthropic's new Claude Mythos 5.2 appeared in some users' model pickers. One early tester turned it loose on their own websites for a full audit of architecture, security, performance, SEO and accessibility, found a long list of problems, and spent about 12 hours letting it rebuild and patch things. The obvious next experiment is to run Neon on the same unpatched code and see if it finds the same holes. That's the new normal: frontier models auditing your systems overnight.
The safety numbers
Because Neon brings frontier capability to a cheaper, faster model, misuse is a fair worry. Google says it kept Carbon-level safeguards against cyber and biosecurity misuse while refusing far less legitimate work. Among the numbers it reported:
- Refused just 1.98% of tasks on Terminal-Bench 4.0 and 0% on Terminal-Bench Science.
- 0.0% cheating on Terminal-Bench 4.0, and shortcuts on only 2.45% of tasks in Google's coding-cheating test, the lowest of any Gemini 4 model.
- A 0.09% misaligned action rate on its agentic safety suite and a 99.2% pass rate on an agent honesty test.
- Its lowest fabrication rate yet in internal Antigravity testing.
- With real-time safeguards at its highest thinking level, it resisted every attack in Gray Swan's indirect prompt injection benchmark.
These are Google's own evaluations, so read them as claims. But they target exactly the problems that have dogged rival launches this month, including high hallucination rates.
The bot arena: a different kind of test
Benchmarks are one thing. Watching models compete is another, and that's where one of the most interesting early tests of Neon is happening.
Victris Arena is a programming game for people and AI agents. You, or your coding agent, write a bot. The bot connects to the arena over a WebSocket and fights other bots in fast one-on-one matches on a 21-by-21 tile grid. Each tick it receives the battlefield as JSON and answers with a move and an attack within a second.
There's real depth for something so simple:
- A main shot that can bank off a wall, which turns aiming into a geometry puzzle.
- Abilities with cooldowns, including an area spell and a point-blank shove.
- Fog of war, so bots lose sight of each other behind pillars.
- Dashes that make you invulnerable for a tick, so defense is all about movement.
- Matches are deterministic, every game has a replay, and a public ladder rates every bot version.
Coding agents such as Codex or Claude Code can connect through the arena's MCP server, write a bot, test it in an offline simulator and enter it in rated matches. The site even keeps agent standings comparing which AI tools build the best bots.
In one tester's setup, a bot built by GPT-6.2 Astra the day before reached a rating around 2,400, roughly 14th on the leaderboard. Within four or five hours of Neon's release, a Neon-built bot running at high reasoning (not even the maximum) was starting to creep past it, and appeared to be learning faster than the other models.
There was a twist. The other models, including Fable 5.2, were sharing notes: each dropped its insights and bot versions into a shared folder, and a "council" skill had them discuss and improve together. Neon wasn't even connected to that shared knowledge yet. With access to it, the tester expected Neon's next version to beat everything, including Fable. That's a prediction, not a result. But it's the kind of hands-on signal that benchmarks miss.
Why bot battles matter more than they look
This isn't just a game. It connects to a serious line of research.
Earlier this year, Tokyo's Sakana AI published a paper called Digital Red Queen: Adversarial Program Evolution in Core War with LLMs. Core War is a programming game from 1984 in which small programs called "warriors," written in an assembly-like language called Redcode, battle for control of a virtual computer.
Sakana's method was deliberately simple:
- Start with one warrior.
- Have a language model evolve a second warrior that beats it.
- Evolve a third that beats the first two, and so on, round after round.
Two things happened. First, warriors became more generally robust, performing better against human-designed warriors they had never seen. Second, and stranger, independent runs that started from different warriors gradually converged on similar behaviors, though not similar code. It's a lot like convergent evolution in biology, where different species evolve the same solution, such as wings or eyes.
Sakana argues that this kind of arms race is a safe sandbox for studying how AI agents might adapt in real adversarial settings like cybersecurity. Many people also see it as a small, contained version of recursive self-improvement: models getting better by repeatedly testing what works against a changing opponent and building on it. An arena like Victris lets anyone watch a version of that loop happen with today's models.
A small story about bots finding bots
How did an early tester even find this arena? The story says a lot about late 2026.
The arena's founder had a bot whose job was to find people online who talk about AI agents and pitch them something relevant. That bot emailed the tester. The tester gets far too much email to see most of it, but their own always-on assistant in Codex flagged it: someone wants to give you a free league to try.
So one person's bot found another person's bot, which decided its human should see the message. Humans still made every real decision: someone set up the outreach bot, the tester said yes, and viewers chose to click. But the filtering layer in between is increasingly machines talking to machines.
Practical takeaways
- If you pay for Google AI Pro or Ultra, select Neon under "Deeper analysis" in the Gemini app. It's likely the best default for serious everyday work.
- For developers, compare cost per finished task, not per-token price. Neon's token prices are higher than some mid-tier rivals, but it may need fewer tokens and fewer retries.
- Watch the November 30 date. Carbon's introductory pricing ends then, which shifts the math in Neon's favor.
- Use the right thinking level. Neon has four, Low to Extra High. Default High for most work; Extra High only for the hardest problems.
- Try planner-executor setups. Neon planning with Hydrogen executing can cut costs on big agent jobs.
- Test coding claims yourself. Run Neon on a real task from your own codebase before switching from Claude.
- Want a fun, hands-on way to learn agents? Bot arenas let your coding agent build, test and improve something measurable with zero risk to real systems.
Counterpoints and risks
- Most numbers come from Google. Several benchmarks are third-party, but Google chose which to highlight.
- Harness effects are huge. A score that triples inside Google's own tool may not carry over to yours.
- Early hands-on results are early. A bot rating from the first few hours, or a private benchmark run on launch day, can change quickly.
- Naming fatigue is real. Four element names, thinking levels and plan tiers make it harder for normal users to know what they're getting.
- Cheaper frontier power cuts both ways. Strong cyber skills at lower cost help defenders, and Google limits the specialist version for a reason.
FAQ
What is Gemini 4 Neon?
Google's mid-tier Gemini 4 model, released November 18, 2026. It's adapted from Carbon and built for complex coding, analysis and enterprise work at higher speed and lower cost per task.
How much does Gemini 4 Neon cost?
$3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens, with a 1-million-token context window.
Is Neon better than Gemini 4 Carbon?
On many benchmarks, it scores comparably or higher at lower cost per task. Carbon remains Google's pick for the hardest, longest projects.
Who can use Neon?
Developers on the paid Gemini API in AI Studio, businesses through Gemini for business, and Google AI Pro and Ultra subscribers in the Gemini app and Antigravity.
Does Neon beat Claude at coding?
Google reports top scores on some coding indexes, but Claude Fable 5.2 and Opus 5.5 have led most coding benchmarks. Independent tests will settle it.
The bottom line
Gemini 4 Neon is Google's clearest sign yet that the AI race is no longer only about the single smartest model. It's about the best intelligence per dollar and per second, and on that measure Neon may be the model to beat right now.
The smart move isn't to believe every chart. It's to put Neon on your real work, in your real tools, and see whether "faster, cheaper and better" survives contact with your own tasks.
Sources
- Introducing Gemini 4 Neon (Google)
- Gemini 4 Neon model card (Google DeepMind)
- Gemini 4 Neon in Google Antigravity (Antigravity blog)
- Google launches Gemini 4 Neon with Carbon quality at Flash speeds (9to5Google)
- Google launches Gemini 4 Neon with Opus 5.5-level coding performance at 25% lower API token prices (VentureBeat)
- Digital Red Queen: Adversarial Program Evolution in Core War with LLMs (Sakana AI)
- About Victris Arena





