Official Gemini 4 Neon model page in Google AI Studio docs, showing the model code gemini-4-neon and a unified 1M-token context window
Official Gemini 4 Neon model page at aistudio.google.com/docs/models/gemini-4-neon, captured 20 November 2026, used as the cover because this comparison productSlug is gemini-4-neon. Image: AppFor.ai screenshot of aistudio.google.com. Image source ↗

The short answer: Neon for faster mid-family work, Carbon for the hardest official reasoning role

If you want a paid Gemini 4 API model for complex software engineering, knowledge work, and science, and you care about a default high thinking level plus an xhigh setting, Gemini 4 Neon is the official mid-family choice, at a flat Standard $3 input and $15 output per million tokens. If you want the model Google calls its strongest reasoning model for tasks that push reasoning depth, including root-cause debugging, architectural design, scientific proofs, and security auditing, Gemini 4 Carbon is the official choice. Carbon defaults to medium and offers only low, medium, and high. Neither model is on the Gemini API Free tier. This comparison does not pick a universal quality winner.

Scope and evidence

This is a source-based comparison of two paid Gemini 4 API models, checked on 20 November 2026: Gemini 4 Neon (API ID gemini-4-neon) and Gemini 4 Carbon (API ID gemini-4-carbon). It uses official model pages, What's new guides, the live Gemini Developer API pricing page, and Google's 18 November Neon announcement. It does not include a recorded same-input test, and it does not treat vendor benchmark or cost-per-task claims as independent scores.

Prices below are official Standard paid-tier rates in USD per 1 million tokens, including thinking tokens in the output price, with implicit caching only. Batch, Flex, and Priority tariff tables were visible as tabs on the pricing page but were not switched and recorded in this session, so those tier prices are not listed. The published Carbon profile on this site still says the model is not yet on sale; the live official pricing and What's new pages checked today are the source for current list rates.

Compare the same criteria

Each line uses official pages for both models as of 20 November 2026. API prices are Standard paid-tier list rates. Google's 32% cost-per-task comparison is a vendor claim from the Neon announcement, not a list-price comparison.

  • Official role. Gemini 4 Neon: official docs place it between Gemini 4 Hydrogen and Gemini 4 Carbon for high-frequency complex tasks, and describe frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows. Gemini 4 Carbon: official docs call it the strongest or most capable Gemini 4 reasoning model, built for sustained agentic workflows, hard architecture and debugging, dense multimodal synthesis, scientific proofs, and security auditing. Carbon, the docs say, spends more thinking tokens per turn to explore and verify its work, which increases per-turn token use and latency.
  • Standard API list price. Gemini 4 Neon: $3.00 input, $15.00 output, and $0.15 implicit context cache per million tokens, with no scheduled change listed on the live pricing page. Gemini 4 Carbon: $2.00 input, $10.00 output, and $0.10 cache through 30 November 2026, then $4.00 input, $20.00 output, and $0.20 cache starting 1 December 2026. Both are paid-tier only. Through 30 November, Carbon's list rates are lower than Neon's. From 1 December, Neon's listed $3 / $15 rates are lower than Carbon's $4 / $20.
  • Thinking controls. Gemini 4 Neon: official levels are low, medium, high (the default), and xhigh. Gemini 4 Carbon: official levels are low, medium (the default), and high. Carbon's official What's new pages do not list xhigh. If you need the official xhigh setting, Neon is the model that documents it.
  • Context, output, and inputs. Both official model pages list a unified 1,048,576-token context window, default synchronous max output of 65,536 tokens, long decode up to 1,048,576 tokens, inputs of text, image, video, audio, and PDF, and text output. Both list implicit caching only, plus Batch, Flex, and Priority consumption options. On these official specs, the two models are aligned.
  • Where you can use them. The live pricing page says neither model is available on the Free tier. Official What's new documentation says Neon is on the paid Gemini API (generateContent and the Interactions API) and with Google AI Pro and Ultra plans in Google AI Studio. Carbon's official What's new page says it is available on the paid tier of the Gemini API and in Google AI Studio. This comparison does not add Gemini app or Antigravity access for Carbon beyond those official pages checked today.
  • Vendor cost-per-task and benchmark claims. Google's Neon announcement, checked again on 20 November 2026, says that at default high, Neon pairs a high Artificial Analysis Intelligence Index score with a 32% lower cost per task than Carbon. That is a cost-per-task claim, not a claim that Neon's $3 / $15 list rates are below Carbon's current promotional $2 / $10 list rates. Treat the 32% figure as Google's published evaluation, not as an independent measurement in this article.

Which fits your real work?

If you are doing complex but frequent coding, document synthesis, or science workflows and you want the official mid-family Gemini 4 model, with a default of high and the option to raise the job to xhigh, Neon is the documented fit. If the job is the hardest official Carbon role (deep root-cause debugging, architectural design, formal proofs, or security auditing) and you want the model Google says spends more thinking tokens to verify its work, Carbon is the documented fit.

If list token price is the main constraint in late November 2026, Carbon's official Standard rates are still $2 / $10 through 30 November, which is below Neon's $3 / $15. From 1 December 2026, that relationship reverses on the same official pricing page. Budget around the date, not around a single frozen number, and remember that Carbon may also use more thinking tokens per turn according to Google's own description.

If you are on the Gemini API Free tier, neither official pricing table lists these models as available. Hydrogen is the Gemini 4 model the same page lists with a Free tier. In that case, neither Neon nor Carbon is a supported free choice.

Cost of getting the work done

On official Standard list prices checked on 20 November 2026, a short-context paid request is cheaper on Carbon through 30 November ($2 input and $10 output per million tokens) than on Neon ($3 and $15). From 1 December, Carbon rises to $4 and $20, while Neon stays at $3 and $15 on the page checked today. Official cache rates follow the same pattern: $0.10 then $0.20 for Carbon, and a flat $0.15 for Neon, implicit only.

Google's 32% lower cost per task for Neon versus Carbon is a vendor evaluation at Neon's default high thinking level. It can diverge from the list-price comparison, especially while Carbon's introductory rates are still in effect, and if Carbon spends more thinking tokens per turn as the official Carbon guide says it does. Use the live list rates for budgeting, and treat cost-per-task claims as the publisher's own evaluation.

Grounding with Google Search or Maps, when you use it, is a paid-tier extra after a shared monthly allowance, then $14 per 1,000 requests, with a documented daily cap on Search. That extra is the same official rule on both model entries. Subscription dollar prices for Google AI Pro or Ultra are not quoted here, because a live official plans page was not used as a price source in this session.

Try both on the same small task

This is a suggested trial, not a recorded test. In a paid Gemini API project, send both models the same repository excerpt or document packet and the same instruction: find the defect or required facts, apply them, and return a structured fix or memo that follows the brief, with no extra filler. On Neon, leave thinking at the official default of high unless you are deliberately testing xhigh. On Carbon, leave thinking at the official default of medium, and run a second pass at high if you want to see the official top Carbon level.

Then compare, without treating the result as a benchmark: whether every required fact or failing test is addressed, whether the answer invents files, numbers, or citations, how many extra tool or edit turns it needed, and what the official usage meter shows for that single run, including thinking tokens and any cache hits.

Decision checklist

Stay on Gemini 4 Neon if you want the official mid-family role, a default high thinking level, or the documented xhigh option, or if you are budgeting against the flat $3 / $15 Standard rates, especially after 1 December 2026. Stay on Gemini 4 Carbon if you want the official hardest-reasoning Gemini 4 role, you are comfortable with a default of medium and a cap at high, or you want the lower introductory $2 / $10 list rates while they last through 30 November 2026.

What would change the choice is a change in official list prices after the November window, a recorded same-input test on your own work, or a need for xhigh, which official Carbon docs do not list. Linked profiles: Gemini 4 Neon and Gemini 4 Carbon. Re-check the live pricing page before you budget a large run.

Sources & further reading

Based on the official documentation below. Features and access can change; check the provider before making a decision.

  1. Official Gemini 4 Neon model page, rendered 20 November 2026 (opens in a new tab)
  2. Official What's new in Gemini 4 Neon, rendered 20 November 2026 (opens in a new tab)
  3. Official Gemini 4 Carbon model page, rendered 20 November 2026 (opens in a new tab)
  4. Official What's new in Gemini 4 Carbon, rendered 20 November 2026 (opens in a new tab)
  5. Official Gemini Developer API pricing page, rendered 20 November 2026 (opens in a new tab)
  6. Google: Introducing Gemini 4 Neon, dated 18 November 2026, checked 20 November 2026 (opens in a new tab)
Something out of date or incorrect?Report a correction

Was this page useful?

Help us make our coverage more useful. This is feedback on the page, not a rating of the tool.

Share this page

Related reading