Gemini 4 Neon By The Numbers: Gains, Costs, And What’s Missing
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Gemini 4 Neon By The Numbers: Gains, Costs, And What’s Missing on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

Google released Gemini 4 Neon on November 18, positioning it between its Carbon and Hydrogen models. Artificial Analysis measured a one-point Intelligence Index gain over Carbon at xhigh, alongside lower cost and shorter task time, but Neon trails the index leaders and some results depend on the effort setting. Google’s comparison table omits those two leaders.

Google released Gemini 4 Neon on November 18, describing it as a model that balances depth and efficiency. Artificial Analysis measured Neon at 61 on its Intelligence Index at xhigh, one point above Gemini 4 Carbon at high, while reporting lower cost and shorter completion time for an index task; Neon still ranked behind two models that Google’s comparison table does not include.

At standard list prices, Artificial Analysis estimates that an index task costs $6.44 with Neon at xhigh, compared with $7.76 for Carbon at high. The measured task took 21.5 minutes with Neon and 37.1 minutes with Carbon. These comparisons use different effort settings. At the same high setting, Neon scored 60, matching Carbon’s high score, for a reported $3.74 per task.

Google lists Neon at $3 per million input tokens and $15 per million output tokens, below Carbon’s standard $4 and $20 rates. Carbon is discounted to $2 and $10 through November 30, according to the source material, so Neon’s listed prices are not cheaper than that temporary offer. Artificial Analysis says Neon generates about 110 output tokens per second, compared with about 55 for Carbon, but uses roughly 15% more output tokens per task, reducing some of the cost advantage.

Results vary by setting. Artificial Analysis reports Neon scores of 61 at xhigh, 60 at high, 55 at medium and 44 at low. The corresponding costs per index task are $6.44, $3.74, $1.90 and $0.75; task times are 21.5, 14.8, 8.3 and 3.7 minutes. Google sets high as the default. The results are measurements from an independent index, not a guarantee of performance or cost for every user workload.

At a glance
reportWhen: Released November 18, 2026; figures acc…
The developmentGoogle released Gemini 4 Neon, a model adapted from Gemini 4 Carbon, with independent measurements showing a small benchmark gain and lower cost and task time.

Gemini 4 Neon

Four models in seven weeks

Google’s Gemini 4 releases side by side, each at its best tested setting in the independent Artificial Analysis index. The best value in each row is outlined.

ArgonCarbonHydro.Neon ReleasedSep 30Oct 28Nov 12Nov 18 Index50605461 Cost/task$4.97$7.76$3.16$6.44 Min./task29.737.115.321.5 Output $$10$20$7.50$15 Carbon $2/$10 discount ends Nov 30Neon: +1 point, $1.32 less per task
  1. Four models in seven weeksArgon on September 30, Carbon on October 28, Hydrogen on November 12, Neon on November 18. Neon is adapted from Carbon.
  2. A small step in the indexNeon at xhigh scores 61, one point above Carbon at high. It ranks third among model families, behind GPT-6.2 Astra and Claude Fable 5.2.
  3. A bigger step in cost and timeOne index task costs $6.44 instead of $7.76 and takes 21.5 instead of 37.1 minutes. Neon’s tokens are 25% cheaper than Carbon’s standard price.
  4. What to watchUntil November 30, Carbon still sells at a discounted $2/$10. Google’s own table compares Neon with Claude Opus 5.5 and GPT-6.2 Sol, not with the index leaders.

Source: Artificial Analysis, Gemini 4 Neon model page and article, accessed Nov 19, 2026 (Intelligence Index v5.0.1, cost and time per task at standard prices, release dates); Google, “Introducing Gemini 4 Neon”, Nov 18, 2026; Gemini 4 Neon model card. Output price in US$ per million tokens, standard rates. Redrawn by Thorsten Meyer AI.

Graphic: Powered by Thorsten Meyer AI

Lower Task Costs, Setting by Setting

The release presents a trade-off for teams choosing between benchmark performance, latency and spending. In Artificial Analysis’s figures, moving Neon from high to xhigh adds one index point but increases the measured task cost from $3.74 to $6.44. That makes the selected effort level a substantial part of the result, rather than a minor configuration detail.

For coding-focused users, the same source reports a near tie on its Coding Agent Index: Neon scored 58.6, Claude Fable 5.2 scored 58.5, and Claude Opus 5.5 scored 58.2. The reported Neon task cost was $6.86, against $15.47 for Fable and $12.61 for Opus. These figures suggest a possible cost advantage on that test, but they do not establish equivalent results across other coding tasks or production use.

Roborock Qrevo S Pro Robot Vacuum and Mop with Multifunctional Dock

Roborock Qrevo S Pro Robot Vacuum and Mop with Multifunctional Dock

Shark Rocket Pet Plus Corded Stick Vacuum, Bordeaux/Silver, HV322

Shark Rocket Pet Plus Corded Stick Vacuum, Bordeaux/Silver, HV322

As an affiliate, we earn on qualifying purchases.

A Fourth Gemini Model in Seven Weeks

Neon is Google’s fourth Gemini 4 model in about seven weeks, following Argon on September 30, Carbon on October 28 and Hydrogen on November 12. Google describes Carbon as suited to difficult, long-running projects and Hydrogen to fast everyday execution, with Neon between them. Its model card says Neon is adapted from Carbon; for details on architecture, training data and hardware, it refers readers to Carbon’s documentation.

Google says Neon accepts text, images, audio and video, produces text, and has a one-million-token context window. Its knowledge cutoff is March 2026. Output is up to 64,000 tokens by default, with up to one million available through background long decode, according to the model card. Google lists access through the Gemini API in AI Studio and several paid business, developer and consumer products. A cyber-defense version is being made available to eligible members of Google’s Fairwind Program, with access expected over the following few days.

“frontier-level intelligence with fast end-to-end task completion.”

— Google, in its launch post

Benchmark Comparisons Still Incomplete

Google’s evaluation document reports Neon ahead in 16 of 19 rows, but its comparison set includes Claude Opus 5.5 and GPT-6.2 Sol, not Claude Fable 5.2 or GPT-6.2 Astra, the two models ahead of Neon in Artificial Analysis’s family rankings. On Google’s own table, Opus leads Neon on Terminal-bench 4.1 and FrontierCode 2.0, while GPT-6.2 Sol leads on Chartography. The different test sets and methodologies mean results should not be treated as directly interchangeable.

Artificial Analysis also records mixed results: Neon at xhigh scores 63.6% on Terminal-Bench 4.0, below Carbon at high with 65.2%. Its AA-Omniscience hallucination rate is 14.4%, compared with Carbon’s 11.4%. Google separately claims Neon had its lowest fabrication rate yet in internal Antigravity testing, but the source material says that measure is not independently checkable. The full basis for comparing those internal results with public benchmarks is not provided.

Watch for Broader Independent Tests

The next useful evidence will be independent results across more tasks, especially direct tests against GPT-6.2 Astra and Claude Fable 5.2 using clearly matched settings. Further real-world cost and latency measurements could also show whether the index-task advantage holds for different prompts, context lengths and workloads.

For now, users and developers can compare Neon’s effort levels and published prices against their own requirements. The source material gives no date for additional Google benchmark disclosures or for a wider release of the cyber-defense version beyond eligible Fairwind members’ access over the following few days.

Key Questions

How much better did Gemini 4 Neon score than Carbon?

Artificial Analysis scored Neon at xhigh at 61 and Carbon at high at 60 on its Intelligence Index. At the same high setting, Neon also scored 60. These are benchmark results, not a measure of every use case.

Is Neon cheaper than Carbon?

Neon’s standard listed rates are $3 per million input tokens and $15 per million output tokens, compared with Carbon’s standard $4 and $20. The source says Carbon is temporarily discounted to $2 and $10 through November 30, so Carbon’s offer is lower during that period.

Which models rank ahead of Neon?

In the Artificial Analysis model-family ranking described in the source, GPT-6.2 Astra and Claude Fable 5.2 rank ahead of Neon. Google’s own comparison table does not include either of them.

What remains uncertain about Google’s performance claims?

Google’s claim about its lowest fabrication rate comes from internal Antigravity testing, which the source says outsiders cannot check. Independent tests also show mixed results, and the available comparisons use different benchmarks and model settings.

Source: ThorstenMeyerAI.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
You May Also Like

Six Patterns That Make AI Writing A Chore To Read—and How To Help

A preprint comparing 14 language models with human writing identifies six patterns linked to harder-to-process sentences, but does not test reader comprehension.

Ellison’s Universal High Income Meets Real-World Income Trials

Larry Ellison proposed taxing value added by robots to fund universal high income. South Korea’s rural cash pilot is already testing a smaller model.

OpenAI Dev Day 2026: Breaking Down The Key News

OpenAI used DevDay on Sept 29, 2026 to relaunch as an agent platform: Dots, a plugin ecosystem, Decisions API and GPT-6.1 Sol headlined 20+ announcements.

Revolutionize AI Analysis With OlmoEarth Custom Embeddings

OlmoEarth Studio now supports on-demand creation and export of tailored satellite data embeddings, enhancing Earth observation analysis capabilities.