Google released Gemini 4 Neon on 18 November, the middle model of its Gemini 4 family between Hydrogen and Carbon. On the independent Artificial Analysis Intelligence Index it matches Carbon at the high thinking level for less than half the cost per task at list prices. For European organisations the score matters less than three questions it raises: when a fair comparison is possible, who sets the thinking level, and what changed in the Gemini tools your staff already use.

What Gemini 4 Neon is and where it sits

According to Google’s announcement, Neon is built for complex coding, analytical reasoning and enterprise workflows where depth and speed both count. It is adapted from Gemini 4 Carbon, has a 1 million token context window and reads text, images, audio, video and PDFs. It answers in text only, and its knowledge runs to March 2026.

Artificial Analysis calls Neon Google’s fourth new model in seven weeks. The family now looks like this.

Model Role Price per million tokens, input and output In the Gemini app
Hydrogen fast everyday work and subagents $1.50 and $7.50 free via Auto, selectable from AI Plus
Neon deeper analysis, code and knowledge work $3 and $15 AI Pro and AI Ultra
Carbon the hardest, longest tasks $2 and $10 until 30 November, then $4 and $20 AI Ultra only
Argon the first Gemini 4 model, for code and analysis $2 and $10 until 30 November, then $4 and $20 removed, Neon takes over

We covered Gemini 4 Carbon and its cyber-first rollout and Gemini 4 Hydrogen as the execution layer for agents earlier. Neon now sits between them.

What the independent numbers show

Artificial Analysis measured Neon at 61 on its Intelligence Index at the xhigh setting, one point above Carbon at high and third overall, behind GPT-6.2 Astra and Claude Fable 5.2 at their maximum settings. At high, Neon’s default, it scores 60, the same as Carbon.

It is also faster. Neon writes about 110 tokens per second against roughly 57 for Carbon, and finishes an average index task in about 21 minutes instead of 36. In coding agent work through Google Antigravity it scores 59 on the Coding Agent Index, against 58 for Claude Opus 5.5 and Claude Fable 5.2 in Claude Code, at $6.86 per task against $12.61 and $15.47.

Google’s own announcement lists first places in legal research, tax work and office automation. Part of that can be checked. On the Vals Index of professional tasks, Neon ranks second of 42 models at 70.9%, just ahead of Carbon. Most of the other benchmarks in the launch post are Google-reported.

Why a fair price comparison has to wait until December

Per token, Neon is not cheap. It costs $3 per million input tokens and $15 per million output tokens. Carbon and Argon cost $2 and $10 until 30 November as an introductory price, then $4 and $20 from 1 December. Until the end of the month Neon is 50% more expensive per token than Carbon. After that it is 25% cheaper.

Neon does not use fewer tokens either. Artificial Analysis found it writes about 15% more output tokens per task than Carbon, so the whole saving comes from the lower token price. At list prices a task at high costs $3.74 with Neon and $7.76 with Carbon.

Convert that measurement to Carbon’s introductory price and Carbon drops to about $3.90 per task. Until 30 November the two are roughly level. A Data Studios calculation for a 40-turn coding agent even puts Carbon at $2.12 against $3.19 for Neon before that date. Any pilot that compares the two this month measures a promotion, not a running cost.

Bar chart that flips: until 30 November Neon and Carbon cost about the same per task, from 1 December Carbon doubles
Same score, different bill. Neon’s advantage is entirely in its token price, and Carbon’s doubles on 1 December. Source: Artificial Analysis, converted to Carbon’s introductory price.

The thinking level is now the budget lever

Neon has four thinking levels: low, medium, high and xhigh. The higher the level, the longer the model reasons before it answers and the more tokens you pay for. Artificial Analysis measured each level separately.

Thinking level Index score Cost per task Time per task
low 44 $0.75 3.7 minutes
medium 55 $1.90 8.3 minutes
high 60 $3.74 14.8 minutes
xhigh 61 $6.44 21.5 minutes

Going from medium to high adds five points and doubles the cost. Going from high to xhigh adds one point for 72% more money and 45% more waiting. Between the cheapest and the most expensive setting of the same model, the cost per task differs by a factor of almost nine.

That makes the thinking level a budget decision, and the defaults work against you. In the API Neon defaults to high, where Carbon defaulted to medium, so a team that only swaps the model ID will pay for more reasoning without asking for it. Set thinking_level explicitly per workload, reserve xhigh for the tasks where one point matters, and put it in your cost reporting next to the model name. We saw the same pattern with Claude Fable 5.2’s effort levels.

Curve of index score against cost per task for Gemini 4 Neon's four thinking levels, flattening after high
After high the curve flattens. The last index point costs more than every step before it combined. Source: Artificial Analysis, 18 November 2026.

Your staff’s Gemini changed overnight

Anyone with a Google AI Pro subscription now sees Neon in the Gemini app’s model picker, next to Auto and Hydrogen. Argon and the older 3.1 Pro are gone. Carbon stays on AI Ultra. Free accounts and Google AI Plus do not get Neon, and the rollout is gradual.

Two details matter for organisations. First, when users pick a model manually without choosing a thinking level, Gemini sets it to Low, so the same prompt can give a shallower answer than people got from Argon last week. Second, Gemini Tasks, the mode in which Gemini works through multi-step jobs, can now run on Neon. If employees use personal AI Pro accounts for work, the model behind their output changed without anyone in IT deciding it. It is a good moment to check which accounts are in use.

Where your data goes

Neon is available in the Gemini API through Google AI Studio, in Gemini for business, in Antigravity for paying AI Pro and AI Ultra subscribers, and on the Gemini Enterprise Agent Platform, formerly Vertex AI. For processing that stays in Europe, the Agent Platform is the route, because it offers multi-region and regional endpoints.

gemini-4-neon already appears in the Agent Platform’s endpoint tables and its data residency page, which Google updated on 19 November. Check the EU multi-region column for your project before you move personal data to it. At its launch, Hydrogen received Provisioned Throughput in the eu multi-region, which is the precedent to ask your Google account team about.

The model card and the AI Act

Under the EU AI Act, providers of general-purpose AI models have had to supply documentation to downstream providers since August 2025, and Google signed the General-Purpose AI Code of Practice in 2025. For a deployer, the Neon model card is therefore the first document to file. It states that Neon is adapted from Carbon and that its knowledge stops in March 2026.

Google reports that Neon refuses fewer legitimate tasks than Carbon with comparable safeguards against cyber and biosecurity misuse, and that no attack succeeded on Gray Swan’s indirect prompt injection benchmark at its highest thinking level. Those are vendor tests. A separate Gemini 4 Neon for cyber goes only to vetted defenders through Google’s Fairwind Program. For the wider timeline, see our overview of the EU AI Act after the Digital Omnibus.

Where Neon falls short

  • Terminal work. On Vals AI’s Terminal-Bench 4.0, Carbon scores 63.1% against 61.6% for Neon. For agents that mostly work in a shell, Carbon is still the stronger choice.
  • Verbosity. Neon writes more tokens than Carbon. The token price hides that, your invoice does not, especially at xhigh.
  • Latency at the top levels. At high it takes about 41 seconds before the first word of the answer appears, at xhigh about 56 seconds. Customer-facing chat needs medium or lower.
  • Mostly vendor numbers. Only part of the launch benchmarks is independently measured. When Argon launched, Bloomberg reported that some Google staff doubted whether strong scores held up in daily work.
  • No free API tier. Batch and flex processing halve the price to $1.50 and $7.50, but every test costs money.

What European teams should do now

  1. Re-run your evaluations after 1 December at list prices, on your own tasks, per thinking level.
  2. Set thinking_level explicitly in every integration and cap xhigh to named use cases.
  3. Tell Gemini app users about the Low default and which level to use for documents and analysis.
  4. Check the EU endpoint on the Agent Platform before any personal data reaches Neon.
  5. Update your model register with the model ID, the date you switched, the thinking level and the model card version.
  6. Set budget alerts before pilots start, because there is no free tier to absorb mistakes.

Frequently asked questions about Gemini 4 Neon

Is Gemini 4 Neon available in the EU?

Yes. It is in the paid Gemini API, in the Gemini app for AI Pro and AI Ultra subscribers and on the Gemini Enterprise Agent Platform. For EU-only processing, check the EU multi-region on the Agent Platform for your project.

How does Gemini 4 Neon compare with Gemini 4 Carbon?

Both score 60 at high on the Artificial Analysis index. Neon is about twice as fast and, from 1 December, costs less than half as much per task. Carbon remains stronger in terminal work and is meant for the longest, hardest jobs.

Does Neon replace Gemini 4 Argon?

In the Gemini app, yes. AI Pro users get Neon instead of Argon. Google has not announced an end date for Argon in the API.

Is there a free tier for Gemini 4 Neon?

No. The API starts at $3 per million input tokens and $15 per million output tokens, or half that in batch and flex mode.

The model will change again, your evaluation routine should not

Four Gemini models in seven weeks, each with four thinking levels and a price that changes on a fixed date, is not a choice you make once. The organisations that benefit from Neon are the ones with a repeatable test of their own tasks, a cost per task they track over time and context they can move between models. That is the argument we made in models are rented, context is owned, and Neon is the next model that proves it.