Gemini 4 Neon is Google's new AI model, released on 18 November 2026 in the Gemini API, Gemini for business and the Gemini app. It sits between Gemini 4 Carbon and Gemini 4 Hydrogen, costs USD 3 in and USD 15 out per million tokens, and can read 1 million tokens in one request. Artificial Analysis gives it 61 points, one behind GPT-6.2 Astra, but also measures it as slow and verbose per task.
What is Gemini 4 Neon?
According to Google, Gemini 4 Neon is built for fast, finished task completion in software engineering, knowledge work and scientific workflows. Google points to spreadsheets, PDFs, charts, audio and video as typical business work, and says Neon can pull all of it into one 1 million-token context window and turn it into finished deliverables. The model accepts text, image, speech and video, and replies with text (Artificial Analysis).
According to Google, Neon is available like this:
- Gemini API in Google AI Studio for developers, on the paid tier (model ID
gemini-4-neon). - Gemini for business, through the API and in the model picker.
- Google Antigravity for paying Google AI Pro and Ultra subscribers.
- The Gemini app for Pro and Ultra subscribers.
Neon is not available on the free tier of the API (Google's pricing page). A separate variant, Neon for cyber, is rolling out to trusted defenders in Google's Fairwind Program.
What is new compared with Gemini 4 Carbon and Hydrogen?
Google places Neon between two other models in the Gemini 4 family. Carbon is for the hardest, longest-running projects, and Hydrogen for fast everyday execution. Neon is meant to balance depth and efficiency: top-level reasoning for demanding work that repeats often, faster and cheaper than the largest models. Google's pricing page describes Carbon as its most capable reasoning model and Hydrogen as the model for everyday coding and agentic workflows.
Four differences matter most for a business, according to Google:
- Lower cost per task than Carbon: at the default high setting, Google reports a 32 percent lower cost per task than Carbon on Artificial Analysis. The figure is Google's own and refers to the high setting.
- Four reasoning levels: the default is high, and xhigh gives deeper reasoning for the hardest tasks. You can tune how much intelligence and token spend each task gets.
- Built for long agent sessions: Neon is meant to carry tool-heavy tasks end to end, reread fewer files and handle context compaction. In Antigravity it can direct several Hydrogen agents in parallel.
- The same safety level as Carbon: Google says Neon keeps Carbon-level protections against misuse and prompt injection, with the lowest shortcut rate (2.45 percent) of any Gemini 4 model in Google's own coding test. These figures are Google's.
The price is not lower than Carbon's right now. Carbon costs USD 2 in and USD 10 out through 30 November 2026, and USD 4 and 20 from 1 December (Google's pricing page). At that point Neon becomes 25 percent cheaper per token. Neon costs twice as much as Hydrogen (USD 1.50 and 7.50).
Benchmarks: how good is Gemini 4 Neon?
The table shows the tests Google highlights at launch. Where an independent source has confirmed the figure, it says so. The rest is "according to Google" and has not been verified. What the tests measure is explained in AI benchmarks explained.
| Test | Neon | What the test measures |
|---|---|---|
| Artificial Analysis Intelligence Index v5 (xhigh) | 61 points (Google and Artificial Analysis) | Composite score across nine tests in knowledge work, coding and science |
| Vals Index | 70.9%, #2 (Google and Vals) | Composite score across finance, legal work and coding, among others |
| Legal Research Bench | #1 (Google and Vals) | Legal research |
| Tax Agent Bench | #1 (Google and Vals) | Tax tasks solved by an AI agent |
| Artificial Analysis Coding Agent Index | 58.6, #1 (according to Google) | Coding agents working on their own |
| Terminal-Bench 4.1 | 65.2% (according to Google) | Tasks solved in a terminal |
| Agents' Last Exam v1.1 | 75.8% (according to Google) | Multi-step work tasks of the kind a professional does |
| HLE-Diamond | 71.3% (according to Google) | Very hard expert-level questions |
Source: Google, Introducing Gemini 4 Neon, 18 November 2026, Vals AI and Artificial Analysis, read on 19 November 2026.
Google also says Neon is best in class on Terminal-Bench Science (70.5%) and RiemannBench (84.0%), and #2 on Arena's new Agent Arena Index (1336 points). Vals has measured Terminal-Bench-Science itself and got 62.86 percent, #2 of 48, so the two figures cannot be compared directly.
What independent measurements say
Artificial Analysis gives Neon 61 points on Intelligence Index v5 at its highest setting (xhigh). That is one point behind GPT-6.2 Astra (62) and four behind Claude Fable 5.2 (65). Google describes the placement as #3 overall.
| Model (setting) | Intelligence Index v5 | Cost per task | Time per task |
|---|---|---|---|
| Claude Fable 5.2 (max) | 65 | USD 26.54 | 41.8 min |
| GPT-6.2 Astra (max) | 62 | USD 4.00 | 13.9 min |
| Gemini 4 Neon (xhigh) | 61 | USD 6.44 | 21.5 min |
| Gemini 4 Carbon (high) | 60 | USD 7.76 | 37.1 min |
| Gemini 4 Hydrogen (xhigh) | 54 | USD 3.16 | 15.3 min |
Source: Artificial Analysis, LLM Leaderboard, read on 19 November 2026.
Vals AI, which tests models on industry tasks, ranks Neon #2 of 42 on the Vals Index with 70.87 percent, behind Claude Fable 5.2 (71.31) and ahead of GPT-6.2 Astra (67.66). Neon is #1 on Legal Research Bench (56.25%) and Tax Agent Bench (54.89%), and #1 on the Vals Finance Index and Vals Legal Index. The results suit businesses that work with documents, accounting and legal matters.
The independent measurements also carry two caveats:
- Slow per task. Artificial Analysis spends 21.5 minutes per task on Neon, against a median of 12.1 minutes for comparable models. Vals reports 31 minutes and 25 seconds per test.
- Verbose. Neon wrote 283 million tokens in Artificial Analysis' test, against a median of 65.1 million. That is why a task costs USD 6.44 even though the token price is low, and Artificial Analysis calls the model somewhat expensive compared with models in the same price range.
Artificial Analysis measured Neon at xhigh, while Google gives high as the default setting. The figures therefore do not refer to the same setting, and cost per task is lower at lower settings.
Gemini 4 Neon or GPT-6.2 Astra: what is the difference?
Neon is cheaper per token but dearer per task in Artificial Analysis' measurement, because it writes more. On the Vals Index, Neon is ahead of Astra. How Astra is built and priced is covered in GPT-6.2 Astra and Sol.
| Gemini 4 Neon | GPT-6.2 Astra | |
|---|---|---|
| Price in / out per million tokens | USD 3 / 15 | USD 8 / 40 |
| Cached input per million tokens | USD 0.15 | USD 0.40 |
| Context | 1 million tokens | Up to 922,000 tokens in |
| Intelligence Index v5 (Artificial Analysis) | 61 points | 62 points |
| Cost per task (Artificial Analysis) | USD 6.44 | USD 4.00 |
| Time per task (Artificial Analysis) | 21.5 min | 13.9 min |
| Vals Index | 70.87% | 67.66% |
| Agents' Last Exam v1.1 (each company's own figures) | 75.8% (Google) | 78.3% (OpenAI) |
Source: Google, OpenAI, Artificial Analysis and Vals AI, read on 18 and 19 November 2026. Prices as of 18 November 2026.
Astra costs 2.7 times as much as Neon per token, both in and out. Even so, Astra is cheaper per task at Artificial Analysis (USD 4.00 against 6.44) and faster. The two companies measured Agents' Last Exam separately, so the figures do not show a shared ranking. Try both on your own tasks before you choose.
How much does Gemini 4 Neon cost?
In the Gemini API you pay per token, and the prices apply from the launch on 18 November 2026. Neon is not on the free tier. The tokens the model thinks with count as output and are billed.
| Model | Input | Cached input | Output | Context |
|---|---|---|---|---|
| Gemini 4 Neon | USD 3 | USD 0.15 | USD 15 | 1 million tokens |
| Gemini 4 Hydrogen | USD 1.50 | USD 0.075 | USD 7.50 | 1 million tokens |
| Gemini 4 Carbon (through 30 Nov 2026) | USD 2 | USD 0.10 | USD 10 | 1 million tokens |
| Gemini 4 Carbon (from 1 Dec 2026) | USD 4 | USD 0.20 | USD 20 | 1 million tokens |
| GPT-6.2 Astra | USD 8 | USD 0.40 | USD 40 | Up to 922,000 tokens in |
Source: Google's Gemini API pricing page and Google's model page for Carbon, read on 19 November 2026. Price per million tokens in USD. The Astra figures are from GPT-6.2 Astra and Sol.
The table below shows a made-up job of 1,000 answers where each answer uses 1,000 input tokens and 500 output tokens, with no caching:
| Model | Calculation | 1,000 answers |
|---|---|---|
| Gemini 4 Hydrogen | 1.50 + 0.5 × 7.50 | USD 5.25 |
| Gemini 4 Neon | 3 + 0.5 × 15 | USD 10.50 |
| Gemini 4 Carbon (from 1 Dec) | 4 + 0.5 × 20 | USD 14 |
| GPT-6.2 Astra | 8 + 0.5 × 40 | USD 28 |
Three things affect the bill:
- Cached input. Cached input costs USD 0.15, 95 percent below the regular input price. Google's pricing page says only implicit caching is supported.
- Batch. The batch price for Neon is USD 1.50 in and USD 7.50 out, half the standard price.
- Tools. Grounding with Google Search includes 5,000 free requests per month, shared across all Gemini 3 models and newer, then USD 14 per 1,000 requests.
In the Gemini app, Neon is available to Google AI Pro and Ultra subscribers. Google's launch post gives no subscription price.
Should your business use Gemini 4 Neon?
Neon fits best where you have large amounts of material and want finished deliverables: long documents, spreadsheets, PDFs, code and agent work. The Vals measurements on finance, legal and tax point the same way. For simple high-volume tasks, Hydrogen is cheaper. If response time matters, expect Neon at xhigh to be slow per task.
| Task | Model to test first |
|---|---|
| Sorting and short answers in large volumes | Gemini 4 Hydrogen |
| Long documents, spreadsheets, legal and accounting work | Gemini 4 Neon |
| Large, long-running and demanding projects | Test Neon against Gemini 4 Carbon and GPT-6.2 Astra on your own examples |
Privacy and data. Google's pricing page states that data from the paid tier of the Gemini API is not used to improve Google's products. We have not found a source saying that Neon processes data in the EU or EEA, so do not assume it until it is in Google's terms and data processing agreement. If you handle personal data, settle this first, see AI and GDPR in Norway.
How to try the model. Neon is in Gemini for business, through the API and in the model picker. Developers can test it in Google AI Studio on the paid tier. Pick 10 to 20 real tasks, for example a long spreadsheet or a contract, and compare answers, time and cost with the model you use today.
If you want help choosing and setting up the right model for your business, AIKI can go through your workflow with you and build the solutions, see AI development. More models are compared in Qwen4 Max.
Sources
- Google: Introducing Gemini 4 Neon, 18 November 2026 (availability, price, context window, benchmarks, place in the Gemini 4 family, safety)
- Google: Gemini API pricing and model page for Gemini 4 Carbon (price for Neon, Hydrogen and Carbon, context window, batch, data terms)
- Artificial Analysis: Gemini 4 Neon and LLM Leaderboard (Intelligence Index v5, cost and time per task)
- Vals AI: Gemini 4 Neon and front page (Vals Index, Legal Research Bench, Tax Agent Bench)
- 9to5Google: Gemini 4 Neon launches, 18 November 2026 (repeats Google's figures)
- OpenAI figures for the comparison are taken from GPT-6.2 Astra and Sol.




