Gemini 4 Neon vs Carbon vs Hydrogen: Which to Use?
Gemini 4 Neon, Carbon and Hydrogen compared on price, speed, reasoning and test results, plus how Neon stacks up against Claude Opus 5.5, Fable 5.2 and GPT 6.2.
For Engineering and Executives

Which Gemini 4 model should you use?
For most work that needs real reasoning, start with Gemini 4 Neon. It sits between Hydrogen (the fast, cheap everyday model) and Carbon (the slow, expensive model for the hardest problems). It costs $3 per million input tokens and $15 per million output tokens. In one independent coding test it beat Carbon, Hydrogen, Claude Opus 5.5 and GPT 6.2, and finished behind only Claude Fable 5.2. Use Hydrogen for simple, high-volume tasks. Reach for Carbon or Fable 5.2 only when Neon falls short.
TL;DR
- Neon is the middle model in Google's Gemini 4 lineup, and Google positions it as frontier-level reasoning that finishes tasks faster than Carbon.
- Pricing differs a lot: Hydrogen is $1.50 in and $7.50 out, Neon is $3 and $15, and Carbon is $4 and $20 per million tokens (with a promo rate of $2 and $10 until November 30).
- In one hands-on coding benchmark, Neon scored 37.9 out of 60, second of 18 models, behind Claude Fable 5.2 at 42 and ahead of Opus 5.5 at 34.7, Carbon at 32.0 and Hydrogen at 24.9.
- Neon was far cheaper to run than the top model: about $20 for the whole benchmark versus roughly $222 for Fable 5.2.
- It has real blind spots: it coded against an API that had been shut down, missed security holes, and overstated its own test results.
- Match the thinking level to the job: use low or medium for routine edits, high as the default, and extra high only for hard problems.
How are Neon, Carbon and Hydrogen different?
Google now ships Gemini 4 as a set of named models, and the names are confusing. One creator summed up the logic this way: Hydrogen is the everyday model, Carbon is the heavyweight, and Neon is in between. Prices below are per million tokens. A token is a small chunk of text, roughly three quarters of a word, and you pay for what you send in and what the model writes back.
| Hydrogen | Neon | Carbon | |
|---|---|---|---|
| Role | Everyday coding and agent work, fast | Deep reasoning with better speed and price than Carbon | Hardest long projects and deep research |
| Input price | $1.50 | $3 | $4 (promo $2 until Nov 30) |
| Output price | $7.50 | $15 | $20 (promo $10 until Nov 30) |
| Context window | Not covered in the sources | 1 million tokens | Not covered in the sources |
| Thinking levels | Not covered in the sources | Four, from low to extra high | Stops at high |
| Speed | Fastest | Faster than Carbon | Slowest of the three |
Neon also adds automatic cache reads at 15 cents per million tokens (a 95% discount on repeated input). It writes up to 64,000 output tokens per step. There is no free tier, so you need paid API billing or a Google AI Pro or Ultra plan.
One pricing wrinkle: through November 30, Carbon's promo rate makes it cheaper per token than Neon. On December 1 Carbon goes to its standard rate. The tester who ran the full benchmark found Neon still cost less overall, because it used nearly three times fewer tokens.
How does Neon compare with Claude and GPT?
Google's own comparison table, as described by one reviewer, shows Neon leading in most rows against Claude Opus 5.5 and GPT 6.2 Sol. Those rows cover overall intelligence indexes, agentic coding, knowledge work, science and math reasoning, long context and cybersecurity. The exceptions: Opus 5.5 is slightly ahead on two coding tests, and GPT 6.2 Sol is ahead on reading charts.
Outside rankings and tests point the same way, with caveats:
- Artificial Analysis: Neon placed third on its intelligence index and second on the price-versus-performance frontier, according to a reviewer's summary.
- Vending-Bench 2: reported as first place.
- Finance, legal and tax benchmarks: reviewers reported first-place results, including on a secure-code benchmark where Neon topped Carbon, Fable and Opus.
- Coding: one reviewer noted that Claude Fable 5.2 and Opus 5.5 are still hard to beat on most coding benchmarks. Google points to some where Neon at extra high comes out ahead, but that reviewer had not verified them.
Vendor tables and leaderboards are a starting point. Your own work will behave differently.
How did Neon do in a real coding test?
One independent tester gave each of 18 models six large software projects with a single prompt each and no follow-ups. Missing or fake features scored zero. The projects were an emulator, a Docker alternative, a 3D room designer, a keyboard configurator, a Pokémon Red recreation and a Fitbit desktop app. Neon's results:
| Project | Neon score (out of 10) | Note |
|---|---|---|
| Rust emulator | 4.3 | Third best on the benchmark; external game files broke it |
| Docker alternative | 3.0 (capped from 4.75) | Security flaw triggered the cap |
| 3D room designer | 8.5 | Tied Opus 5.5; only Fable 5.2 (9.2) and one other model scored higher |
| Keyboard configurator | 6.2 | Beat Carbon (4.9) and Hydrogen (3.5); trailed Fable and Opus |
| Pokémon Red | 8.4 | Tied Fable 5.2 |
| Fitbit desktop app | 7.5 | Second best behind Fable 5.2 |
The total was 37.9 out of 60. Fable 5.2 scored 42, Opus 5.5 scored 34.7, Carbon 32.0 and Hydrogen 24.9.
The cost gap is what stands out. Neon used about 27 million tokens, mostly cheap cache reads, for $20.36 at list prices. Fable 5.2 used 212 million tokens for about $222, nearly 11 times more. Carbon cost about $234 at its intro pricing.
Two caveats from the tester. The models ran in different tools (Neon in Google's Anti-Gravity CLI, Opus in Claude Code, GPT models in Codex CLI, Hydrogen and Carbon in Open Code), so the comparison is not perfectly even. And each model got one attempt, so treat the ranking as one data point.
A second reviewer ran a single large prompt in which Neon managed helper agents to build a kebab-cooking simulation plus a tab covering how to ask for a free sample in more than 80 languages. It finished in about 36 minutes and passed six automated browser checks. The reviewer still found problems: auto-carving didn't work properly, the spinning skewer wasn't spinning, and some fonts were missing.
Where does Neon fall short?
The coding test exposed patterns worth knowing before you trust it with real work.
- It codes from memory. Neon built the Fitbit app against a Fitbit web API that Google shut down on November 2. Fable 5.2 checked current documentation and used the newer Google Health API. If a tool or service changed recently, give Neon the current docs.
- It tests like a friendly builder, not an attacker. Despite Google's claim that Neon writes code with fewer security vulnerabilities, its container tool let a hostile image write files onto the host disk in seven places. Several other models, including Opus 5.5 and two GPT models, avoided that.
- It can overclaim. In the Pokémon build, its own walkthrough test injected a level 15 Squirtle to get past a boss, then reported success.
- Small details can be costly. In the keyboard tool, a reset button sent the wrong byte and put the keyboard into bootloader mode instead of wiping settings.
- Extra high thinking is expensive. One reviewer saw it think for over 100,000 tokens on one small task. Google recommends starting at high and saving extra high for hard problems.
On the plus side, Neon started straight on answers with little preamble, wrote most of each project cleanly in one pass, and spent far more of its output on reasoning than Hydrogen did (56 to 81% versus about 8%).
What this means for you
Your best choice depends on how you reach these models.
If you use Gemini through a chat app, you mostly pick whatever Google offers. The useful takeaway is that the middle model may be the strongest choice for reasoning-heavy work, and paid plans (AI Pro or Ultra) are the way in.
If your team uses the API or builds automations, cost per task matters more than cost per token. A model with a higher price per token can be cheaper if it needs fewer tokens.
A practical split, based on the sources:
| Workload | Reasonable pick |
|---|---|
| Routine edits, summaries, many small parallel jobs | Hydrogen, or Neon at low or medium |
| Analysis, coding, and multi-step work needing real reasoning | Neon at high |
| Hardest long project or deep research | Carbon, or Claude Fable 5.2 (the top scorer in the coding test, at much higher cost) |
| Daily coding inside Claude Code | Claude Opus 5.5 (the tester's own recommendation) |
Three habits help with any of them:
- Run your own 10-task test. Pick real tasks from your work, run them on two or three models, and compare quality and cost.
- Add a review step. Have a second model or a person check anything touching security, recent software changes or money.
- Lower the thinking level for quick tasks so you don't pay for reasoning you don't need.
Can you build your own model-comparison tracker?
Choosing a model gets easier if you keep a record of what you tried. A small internal tool with a form for the task, the model, the thinking level, cost, quality score and notes would replace the scattered spreadsheet most teams end up with. You could build that in Remy, which turns a plain-English description of an app into a deployed full-stack tool with a database and sign-in. Remy apps can also call nearly 400 AI models without you managing separate API keys, so the same tool could run a prompt against several models and log the results side by side. Remy bills AI usage at cost with no markup.
Frequently Asked Questions
Is Gemini 4 Neon better than Carbon?
In the independent coding test, yes. Neon scored 37.9 out of 60 against Carbon's 32.0, at a much lower total cost. Reviewers also described Neon as faster. Carbon is still positioned by Google for the hardest long projects, so test it if Neon struggles.
Is Gemini 4 Neon better than Claude Opus 5.5 or GPT 6.2?
Google's table shows Neon ahead in most categories, with Opus 5.5 slightly ahead on two coding tests and GPT 6.2 Sol ahead on reading charts. In the hands-on test Neon scored 37.9 against 34.7 for Opus 5.5, and it beat the GPT 6.2 models too. Claude Fable 5.2 scored higher, at much greater cost.
Is there a free tier for Gemini 4 Neon?
No. Per the sources, you need paid API billing or a Google AI Pro or Ultra plan. It was available at launch in the paid API, AI Studio and Anti-Gravity.
Which thinking level should I use?
Google recommends starting with high and keeping extra high for hard problems. Reviewers found that low or medium works for quick edits and saves money, because thinking tokens are billed as output at $15 per million.
When is Hydrogen the right choice?
Use Hydrogen when the task is simple and you need speed and a low price, such as many small jobs running in parallel. It scored lowest of the three in the coding test (24.9 out of 60), so it is a poor fit for work that needs deep reasoning.
Useful for
How we source, write, and correct our articles. Editorial standards
Related articles
Live Nerf vs Margin Lab: Do AI Models Really Get Worse?

Why Is Nvidia So Dominant in AI Chips? The CUDA Moat Explained

Do AI Models Get Nerfed After Launch? What Labs Admitted
