# Gemini 4 Neon Launch: $3/$15 Pricing, Benchmarks, and Where It Sits Against Carbon and Hydrogen

> Source: <https://explainx.ai/blog/gemini-4-neon-launch-price-benchmarks-vs-carbon-hydrogen-november-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog>
> Published: 2026-11-19 00:00:00+00:00

**Key takeaway:** Gemini 4 Neon costs $3 per million input tokens and $15 per million output tokens with a 1 million token context window and a $0.15 cached-input rate, and Google reports 61 on the Artificial Analysis Intelligence Index at xhigh.

**Gemini 4 Neon is live.** Google DeepMind released it on November 18, 2026 as the middle model of the Gemini 4 family, priced at **$3 per million input tokens and $15 per million output tokens** with a **1 million token context window**. Google says Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, one point above Carbon. This is the official launch that our earlier [claimed-vs-verified post on the Neon SKUs](https://explainx.ai/blog/gemini-4-neon-vertex-sku-claimed-vs-verified-november-2026) was waiting for.

Below: what Google actually announced, which numbers are Google's own and which come from third parties, how Neon compares with [Carbon](https://explainx.ai/blog/llm-gemini-0-36-gemini-4-carbon-cli-general-access-pricing-2026) and [Hydrogen](https://explainx.ai/blog/gemini-4-hydrogen-vs-argon-carbon-flash-comparison-november-2026), and how to try it today.

## Gemini 4 Neon at a glance

| Question | Answer (source: Google's launch post unless noted) | 
|---|---|
| When did it launch? | November 18, 2026, rolling out that day | 
| Price | $3 input, $15 output, $0.15 cached input per million tokens | 
| Context window | 1 million tokens | 
| Thinking levels | Low, medium, high, xhigh (default is high) | 
| Artificial Analysis Intelligence Index | 61 at xhigh, ranked #3 overall by Google's account | 
| Cost vs Carbon | 32% lower cost per task at the default high level, per Google | 
| Where is it? | Gemini API in AI Studio (paid), Gemini for business, Antigravity (paid Pro and Ultra), Gemini app (Pro and Ultra) | 
| Is there a cyber version? | Yes, Neon for cyber, rolling out to Fairwind Program members in the coming days | 

## What did Google announce?

Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect, framed Neon in the [official launch post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/?ref=explainx) as "frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows."

Google makes five main claims:

1. **Index rank.** Neon scores 61 on the Artificial Analysis Intelligence Index at xhigh, "just behind larger, slower models," and ranks #2 on the Vals Index with 70.9%.
2. **Agent behavior.** It ranks #2 on steerability, meaning how cleanly it adapts to mid-task corrections, on Arena's new Agent Arena Index, with a score of 1336 across more than 4,200 real agent sessions.
3. **Coding.** It leads the Artificial Analysis Coding Agent Index at 58.6, scores 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default high setting.
4. **Cost.** At the default high level Google reports 32% lower cost per task than Carbon, and roughly 50% to 80% lower cost per task than top competing frontier models on the Vals and Agent Arena Pareto charts.
5. **Safety.** It keeps Carbon-level protections against severe misuse and prompt injection, with a 0.09% misaligned action rate on Google's agentic safety suite.

Most of these numbers come from Google's launch post. Several named benchmarks are run by outside groups (Artificial Analysis, Vals, Arena, Zapier, Andon Labs, Gray Swan). Google does not publish raw logs in the post, so treat the ranks as vendor-reported until you see them on the third-party leaderboards. For a wider view of the field, see our [frontier model comparison](https://explainx.ai/blog/frontier-ai-models-compared-fable-5-2-gpt-6-astra-gemini-4-carbon-grok-4-7-kimi-k3-1-deepseek-2026).

## Gemini 4 Neon price: is it cheap?

The headline price is $3 input and $15 output per million tokens. The cached-input price of $0.15 is a 95% discount, which matters a lot for agents that resend the same large context on every turn.

Simon Willison [pointed out](https://simonwillison.net/2026/Nov/18/llm-gemini/?ref=explainx) that the Gemini 4 pricing is confusing right now, and he is right. Google's lineup slide gives Carbon a list price of $4 input and $20 output, but with a promotional $2 and $10 through November 30. Until the promo ends, Carbon is cheaper per token than Neon. After that, the order is Hydrogen, then Neon, then Carbon.

Pricing tiers ladder with the middle step highlighted, showing where Gemini 4 Neon sits between Hydrogen and Carbon

| Model | Input / output per 1M tokens | AA Index (Google slide) | Speed (Google slide, out of 5) | 
|---|---|---|---|
| Carbon | $4 / $20 (promo $2 / $10 to Nov 30) | 60 | 2 | 
| Neon | $3 / $15 | 61 | 3.5 | 
| Hydrogen | $1.50 / $7.50 | 54 | 5 | 

Source: Google's Gemini 4 lineup graphic dated November 18, 2026, as reproduced by Simon Willison. Per-token price is not cost per task. A model that thinks longer can cost more even at a lower rate. Willison ran the same prompt at three thinking levels and recorded output token counts and costs: at high thinking Neon produced 9,691 output tokens for 14.55 cents, against 20,305 tokens and 20.31 cents for Carbon and 5,311 tokens and 3.99 cents for Hydrogen. That is one prompt, not a benchmark, but it shows why Google leads with cost per task.

## Neon vs Carbon vs Hydrogen: which model for which job?

Google's own description is the clearest guide: Carbon for the hardest long-horizon projects and deep research, Neon for deep reasoning across software engineering, knowledge work and science, and Hydrogen for speed and efficiency in everyday coding and agentic workflows.

| Job | Pick | Why | 
|---|---|---|
| Overnight refactor of a big repo | Carbon | Google calls it the model for long-horizon projects | 
| Daily coding agent with tool use | Neon | Four thinking levels, lower cost per task than Carbon per Google | 
| Parallel subagents doing small edits | Hydrogen | Fastest and cheapest, and Google uses it as the subagent in Antigravity | 
| Spreadsheet, PDF and chart analysis | Neon | Google highlights 1M-token multimodal context and enterprise benchmarks | 
| Classification and high-volume extraction | Hydrogen | Per-token price is lowest | 
| Security review and patching | Neon for cyber | Needs Fairwind Program access | 

Our [Hydrogen use-case post](https://explainx.ai/blog/gemini-4-hydrogen-use-cases-demos-search-antigravity-november-2026) covers the cheap end of the family in detail. Argon is missing from Google's lineup slide, and the [Argon-era CWE-bench v2 results](https://explainx.ai/blog/collinear-cwe-bench-v2-100-defensive-patch-tasks-91-cwes-gemini-4-argon-67-8-percent-2026) show why developers will want a replacement for it. Google reports Neon at 70.7% on that benchmark.

## How does Neon work with Antigravity?

Google pitches Neon as a "high-judgment Orchestrator directing parallel Hydrogen subagents" in Google Antigravity. In that setup the larger model plans and reviews while cheaper models do the bulk edits.

An orchestrator shape handing tasks to smaller subagents, representing Gemini 4 Neon directing Hydrogen subagents

Google says Neon cuts redundant file reads, reports progress as it works and handles context compaction reliably in long sessions. On SWE-sweep it reports 12.8% on the mini-SWE harness and 39.0% in Antigravity, which Google presents as proof of what the Antigravity harness unlocks. That is Google's number on Google's harness, so do not expect the same lift in another tool. For earlier Antigravity results see our post on [/teamwork results](https://explainx.ai/blog/google-antigravity-teamwork-open-math-cs-engineering-results-november-2026).

Google's launch post also lists demos built with Neon in Antigravity: a Blender scene, a first-person ski simulator in WebGL2 with about 10,200 lines of shader code, a 3D reconstruction of Alan Turing's 1936 study, and a 100,000-cell Google Sheet that relights an image with over 500,000 formulas. These are Google's showcases, picked by Google.

## Science, enterprise and legal benchmarks

For non-coding work Google reports 70.5% on Terminal-Bench Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond, and calls each best in class. It also reports #1 on the Legal Research Bench and Tax Agent Bench, 56.3% on Zapier's AutomationBench and $16,843 on Andon Labs' Vending-Bench 2, a long-horizon business simulation.

Two cautions. First, benchmark owners often update versions, and Google names versions such as Terminal-Bench 4.1 and Chartography with specific scores that other labs may not report. Second, "best in class" claims compare Neon with models of similar size, as Google defines them. Check the leaderboards for the competitors you care about.

## Neon for cyber: what is claimed

Google says Neon for cyber produces more secure code than Carbon on AutoBaxBench, reaching 76% on the hard tier. It reports 51.26% on WebExploitBench v1.2 at half the cost of competing large models, 90.5% on SEC-bench Pro for finding vulnerabilities, and 70.7% on Collinear's CWE-bench v2 for patching. Access is limited to eligible members of Google's Fairwind Program, who start receiving it over the following days. To request access, Google tells customers to contact their Cloud account team.

This is a defender-gated model, so most developers will use plain Neon. Google says plain Neon carries real-time safeguards against cyber and biosecurity misuse comparable to Carbon's.

## Safety and honesty numbers

Google reports that Neon matched Carbon with 0.0% cheating on Terminal-Bench 4.0 and took shortcuts on 2.45% of tasks on its internal coding cheating evaluation, the lowest of any Gemini 4 model. It also reports a 99.2% pass rate on an agent honesty evaluation and says Neon reduces under-service by 43% to 56% and overreach by 15% to 35% compared with Carbon, depending on effort level. On Gray Swan's indirect prompt injection benchmark, Google says Neon with real-time safeguards at its highest thinking level resisted every attack. The full detail is in the Gemini 4 Neon model card linked from Google's post.

These are internal evaluations. They are useful, but they are not a guarantee for your own agent with your own tools. If you let an agent run code, read email or browse, you still need your own permission and logging layer.

## Neon in the Gemini app

For consumers, Google says Neon is available to Google AI Pro and Ultra subscribers from the model drop-down. It also says Tasks in Gemini can now run on Neon, to work through multi-step jobs "more quickly and efficiently than with larger models" and stretch usage limits further. Google's examples are checking a billing statement for why a utility charge spiked and cross-checking receipts to negotiate a refund. 9to5Google also reported the rollout and the $3/$15 pricing on November 18.

## What this means for what you build or pay

- **Re-test your cost per task, not your per-token price.** Run the same ten real tasks on Hydrogen, Neon and Carbon at high thinking. Record tokens, cost, wall-clock time and pass rate.
- **Use the cached-input price.** At $0.15 per million, a stable system prompt or repository summary costs almost nothing to resend.
- **Watch November 30 and December 1.** Carbon's promo ends, and the price order of the family changes.
- **Do not rely on a model name staying in the lineup.** Argon vanished from the lineup slide. Pin model IDs and read the release notes before migrating.
- **Treat benchmark ranks as leads.** Reproduce the ones that matter for your stack.

## Frequently asked questions

### Does Neon work through the API on day one?

Google says it is rolling out in the Gemini API in Google AI Studio on the paid tier, and also in Android Studio for developers. Rollouts are staged, so a model list may not show it for every project immediately. If it is missing, check again later and read Google's AI Studio changelog rather than guessing.

### Is there a free tier for Neon?

Google's lineup slide lists Hydrogen for everyone and Neon for Google AI Pro and Ultra subscribers and the paid API tier. Google's post does not describe a free API tier for Neon.

### How does Neon compare with models from other labs?

Google compares Neon with "comparable models" and cites 50% to 80% lower cost per task than top competing frontier models on the Pareto charts. Google's launch graphic does include a table against Claude Opus 5.5 and GPT-6.2 Sol, which we break down row by row in our [Neon vs Opus 5.5 vs GPT-6.2 Sol comparison](https://explainx.ai/blog/gemini-4-neon-vs-claude-opus-5-5-vs-gpt-6-2-sol-benchmarks-price-which-model-november-2026). Those are Google-run numbers. Our [frontier model comparison](https://explainx.ai/blog/frontier-ai-models-compared-fable-5-2-gpt-6-astra-gemini-4-carbon-grok-4-7-kimi-k3-1-deepseek-2026) covers the rest of the field.

## Related reading

*Details, prices and rollout status are accurate as of November 19, 2026 and may change. Check Google's launch post and the Gemini API pricing page for current figures.*
