# Gemini 4 Neon vs Carbon: Coding, Speed and Cost

> Compare Neon and Carbon using published coding results, thinking levels, task times and API rates, including Carbon's temporary launch discount.

- Canonical HTML: https://omniakey.com/blog/gemini-4-neon-vs-carbon
- Markdown: https://omniakey.com/blog/gemini-4-neon-vs-carbon.md
- Published: 2026-11-19 (updated 2026-11-19)
- Author: Kiki, COO at OmniaKey
- Language: en
- Category: Comparison
- Reading time: 9 min
- All posts: https://omniakey.com/blog/llms.txt

---

Gemini 4 Neon vs Carbon comes down to two figures from Artificial Analysis's November 18, 2026 run: about 21 minutes per task on Neon against 36 on Carbon, at $6.44 and $7.76. The $7.76 applies Carbon's standard rates, not the $2/$10 introductory price in force through November 30.

- **Gemini 4 Neon** (Faster on tested tasks): xhigh: about 21 minutes and $6.44 per task, index 61.
- **Gemini 4 Carbon** (Tokens at half price until Nov 30): Standard-rate task cost is $7.76, not this month's price; tokens are $2/$10.

Google [published Neon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/) on November 18, 2026 and places it between Carbon, described as the model for the hardest, longest-running projects, and Hydrogen, described as the model for fast everyday execution. Neon costs $3 per million input tokens and $15 per million output tokens. Both models are paid-tier only on the Gemini API, so testing starts with billing enabled. [OmniaKey's Gemini family page](https://omniakey.com/models/gemini) lists the models and current rates.

## Which model fits coding tasks?

Start from the published numbers, because Google's role labels describe intent rather than results. Neon's measured advantage is speed; Carbon's is price, and only through November 30. [Google's launch post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/) puts Carbon on the hardest long-running projects and Hydrogen on fast everyday work, with Neon between them. Those roles are positioning; the differences in the results below run to time and cost.

- **Short edit-and-test loops** are worth testing on Neon: the reported task time is about 40% lower, which is what a latency-sensitive loop needs. That is an aggregate figure, not a promise on every turn.
- **Prompts already working on Carbon** stay there until Neon reproduces the result; Carbon's tokens are half price through November 30, and a regression costs more than the saving.
- **Hardest debugging or architecture work** is an open question rather than a Carbon win: Google assigns that role to Carbon, but the measured differences run to time and cost, not difficulty.

Thinking levels shape that choice too, and the defaults differ. Neon starts at high and goes up to xhigh; Carbon starts at medium and stops at high.

## Gemini 4 Neon vs Carbon at a glance

|Item|Neon|Carbon|
|-|-|-|
|Google's role|Between Carbon and Hydrogen|Hardest, longest tasks|
|Default thinking level|high|medium|
|Highest thinking level|xhigh|high|
|Free tier|none|none|
|Context window|1,048,576|1,048,576|
|Non-streaming output cap|65,536|32,768|

Both models take text, image, video, audio and PDF input and return text ([Neon](https://aistudio.google.com/docs/models/gemini-4-neon), [Carbon](https://ai.google.dev/gemini-api/docs/models/gemini-4-carbon), checked November 19, 2026). The 1,048,576-token window is shared by input, thinking and visible output, so a long prompt leaves less room for the answer. With `background: true`, thinking plus visible output can use up to 1,048,576 minus the input. The cap row covers non-streaming foreground calls ([thinking docs](https://ai.google.dev/gemini-api/docs/thinking)): Carbon stops at 32,768 there, while streaming calls reach 65,536 on both.

## Independent coding results

Artificial Analysis published its [comparison](https://artificialanalysis.ai/articles/gemini-4-neon-intelligence-gain) on November 18, 2026, and the two useful readings were taken at different settings. At high on both models it scored them level on its Intelligence Index. The run behind the time and cost figures put Neon at xhigh and Carbon at high. We have run neither model ourselves.

|Measure|Neon xhigh|Carbon high|
|-|-:|-:|
|Intelligence Index v5.0.1|61|60|
|Reported task time|about 21 minutes|about 36 minutes|
|Task cost, standard rates|$6.44|$7.76|

Task time is the end-to-end time Artificial Analysis reports for its task set. The index behind those rows covers nine evaluations: knowledge work, coding, science and general knowledge, with the task count and repeats on its [method page](https://artificialanalysis.ai/methodology/intelligence-benchmarking). Task cost uses Carbon's standard rates, so it is not what a request costs this month. A separate evaluation, the Coding Agent Index, scored 59 for Neon against 55 for Carbon inside Antigravity's command-line agent. That score folds the agent harness in with the model, so it is not the same evaluation as the rows above, and the two should not be added together.

Neon used about 15% more output tokens on the same tasks. It still cost less per task at standard rates because its output rate is $15 per million against Carbon's $20 rate, so the saving comes from the rate rather than from doing less work.

## How thinking levels affect speed

Both models think by default, at different levels. Neon accepts low, medium, high and xhigh and defaults to high ([release notes](https://aistudio.google.com/docs/whats-new-gemini-4-neon)). Carbon accepts low, medium and high and defaults to medium; minimal is rejected on both ([thinking docs](https://ai.google.dev/gemini-api/docs/thinking)). Google suggests starting at high and reserving xhigh for hard multi-step work, warning that it raises latency and token use and can overthink a simple prompt or exhaust the output cap.

The same named level is a controlled starting point, not matched compute. Google publishes no per-level token budget, and the ceilings differ, so Neon at high and Carbon at high can spend very different amounts of thinking. Artificial Analysis reports about 110 output tokens per second for Neon and about 57 for Carbon in the same comparison.

## Cost during Carbon's launch promotion

Carbon's discount is the reason the token maths and the task maths disagree this month. The [pricing page](https://ai.google.dev/gemini-api/docs/pricing) gives both sets of rates, checked November 19, 2026.

|Standard rate per 1M tokens|Neon|Carbon|
|-|-:|-:|
|Input|$3.00|$2.00 → $4.00|
|Cached input|$0.15|$0.10 → $0.20|
|Output, including thinking|$15.00|$10.00 → $20.00|

Carbon's second figure starts December 1, 2026. Thinking tokens bill at the output rate, and Google also counts internal tool instructions when grounding runs.

One request of 10,000 input tokens and 2,000 output tokens, thinking included, costs $0.06 on Neon. Carbon costs $0.04 through November 30 and $0.08 from December 1. A hundred such requests run $6.00, $4.00 and $8.00. At equal token counts Neon is 50% above the promotional Carbon price and 25% below the December 1 price.

Carbon's measured $7.76 is a standard-rate figure. Halving its rates on the same token counts gives about $3.88, under Neon's $6.44; that is arithmetic on Artificial Analysis's numbers, not a second measurement, and it ignores cache hits and tool fees. So per token Carbon is cheaper this month, per task the standard-rate cost favors Neon, and only the promotional halving puts Carbon below it. From December 1 the token price agrees with the task price.

Both models cache implicitly from 4,096 prompt tokens at 95% off the input rate, and neither offers explicit caching.

If the decision runs past November, budget on the December rates rather than this month's: $4 and $20 per million on Carbon, which is where the token price and the measured task cost stop disagreeing.

## Where each model is available

Neon runs on the paid Gemini API, in AI Studio and Gemini for business, for paid Google AI Pro and Ultra subscribers in Antigravity, and for Pro and Ultra subscribers in the Gemini app ([launch post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/)). Carbon runs on the paid API in AI Studio and Gemini for business, and for Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search ([Carbon post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/)). Neither has a free tier, and Google says paid requests are not used to improve its products.

Both models serve the global endpoint plus the `us` and `eu` multi-regions, so a project that must keep processing inside one of those regions can use either model there ([Cloud docs](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/start/get-started-with-gemini-4?hl=en)).

OmniaKey sells Gemini API access. The scores and times above are published tests, and the rates are Google's official prices. [OmniaKey's Gemini family page](https://omniakey.com/models/gemini) and the [pricing page](https://omniakey.com/pricing) show what a key can call today. No page checked on November 19, 2026 describes a separate NEON subscription plan; the launch names Pro and Ultra.

Carbon's rollout is covered in [Carbon's release and price page](https://omniakey.com/blog/gemini-4-carbon-release-date-price-api), and [Gemini 4 Neon](https://omniakey.com/blog/gemini-4-neon) records how the name surfaced before the November 18 launch. If everyday speed matters more than the top end, [Hydrogen vs Carbon](https://omniakey.com/blog/gemini-4-hydrogen-vs-carbon) covers the cheaper comparison.

## How to compare the two on your own repository

Pick 10 to 20 tasks that currently fail or run slowly. Run both models at the same named thinking level, record pass or fail, total tokens, tool turns and elapsed time, then repeat with the failing tasks at Neon's xhigh. Keep the list close to production traffic, because a set of one-line fixes will not predict a multi-file refactor.

1. Write the task list first, including the cases that stall the current flow: cross-module failing tests, dependency upgrades, a plan that needed a second attempt.
2. Start with Neon high against Carbon high as a controlled baseline, then rerun the failures at Neon xhigh.
3. Compare cost per passing task, not per token. Thinking bills at the output rate, so a cheaper rate with longer thinking can still lose.
4. Recount by November 30 with your own monthly token volume, because Carbon's rates double on December 1.

> Strip&#x20;
>
> `temperature`
>
> ,&#x20;
>
> `top_p`
>
> &#x20;and&#x20;
>
> `top_k`
>
> &#x20;from prompts copied over from Carbon: Neon ignores them and returns 400 for out-of-range values. Set HTTP timeouts to at least 15 minutes for hard tasks, and raise the output cap when you test xhigh so the budget does not truncate the reasoning.

## FAQ

### Is Neon better than Carbon for coding?

At high on both, Artificial Analysis scored them level on its Intelligence Index. With Neon at xhigh and Carbon at high it reported index 61 against 60, about 21 minutes against 36, and $6.44 against $7.76 at standard rates. The Coding Agent Index, a separate evaluation that includes the agent harness, scored 59 against 55.

### Is Carbon cheaper right now?

Per token, yes, through November 30, 2026: $2 input and $10 output per million against Neon's $3 and $15. From December 1 Carbon moves to $4 and $20 a million, which leaves Neon 25% cheaper per token. The $7.76 task cost in the tables is Carbon at standard rates, not this month's price.

### Do the two models default to the same thinking level?

No. Neon defaults to high and accepts up to xhigh; Carbon defaults to medium and stops at high. A matching level name is only a starting condition, because Google publishes no per-level token budget and the ceilings differ.

### Does Gemini 4 Neon have a free tier?

No. Google marks Neon and Carbon as unavailable on the free tier of the Gemini API ([pricing page](https://ai.google.dev/gemini-api/docs/pricing)). The first request needs billing enabled on the project.

### How much output can one request return?

Neon returns up to 65,536 tokens from a non-streaming foreground call, and Carbon 32,768; streaming calls reach 65,536 on both. With `background: true` and a higher cap, thinking plus visible output together can use up to 1,048,576 minus the input, inside the same shared window ([thinking docs](https://ai.google.dev/gemini-api/docs/thinking)).

## Sources checked

- [Google: Introducing Gemini 4 Neon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/) (November 18, 2026)
- [Google: Gemini 4 Carbon now available](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/) (November 10, 2026)
- [Google: Gemini 4 Neon model page](https://aistudio.google.com/docs/models/gemini-4-neon) (checked November 19, 2026)
- [Google: Gemini 4 Carbon model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-carbon) (checked November 19, 2026)
- [Google: Gemini 4 Neon release notes](https://aistudio.google.com/docs/whats-new-gemini-4-neon) (checked November 19, 2026)
- [Google: Gemini API thinking levels](https://ai.google.dev/gemini-api/docs/thinking) (checked November 19, 2026)
- [Google: Gemini API model list](https://ai.google.dev/gemini-api/docs/models) (checked November 19, 2026)
- [Google: Gemini API pricing](https://ai.google.dev/gemini-api/docs/pricing) (checked November 19, 2026)
- [Google Cloud: Get started with Gemini 4](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/start/get-started-with-gemini-4?hl=en) (checked November 19, 2026)
- [Google DeepMind: Gemini 4 Neon model card](https://deepmind.google/models/model-cards/gemini-4-neon/) (November 18, 2026)
- [Google DeepMind: Gemini 4 Carbon model card](https://deepmind.google/models/model-cards/gemini-4-carbon/) (November 10, 2026)
- [Artificial Analysis: Gemini 4 Neon comparison](https://artificialanalysis.ai/articles/gemini-4-neon-intelligence-gain) (November 18, 2026)
- [Artificial Analysis: benchmarking method](https://artificialanalysis.ai/methodology/intelligence-benchmarking) (checked November 19, 2026)
- [Artificial Analysis: about](https://artificialanalysis.ai/about) (checked November 19, 2026)

## Update log

- **November 19, 2026**: first version. Launch, access and rates checked against Google's pages; scores, task time and task cost from Artificial Analysis's November 18, 2026 comparison.
