# Gemini 4 Neon, and more horses in the race

> Vlad Cealicu on Gemini 4 Neon, Google's fourth Gemini 4 model in seven weeks. It scores 61 on the Artificial Analysis Intelligence Index at xhigh, third overall and a point ahead of Gemini 4 Carbon, costs $3 and $15 per million tokens and runs a coding agent task for about half of what Claude Opus 5.5 costs. It loses on Terminal-Bench 4.1, FrontierCode 2.0 and Chartography, and three labs within a few points of each other is good for everyone building agents.

- Canonical: https://www.anchorterminal.com/blog/gemini-4-neon-more-horses-in-the-race
- Markdown: https://www.anchorterminal.com/blog/gemini-4-neon-more-horses-in-the-race.md (~2,300 tokens)
- Slim: https://www.anchorterminal.com/blog/gemini-4-neon-more-horses-in-the-race.min.md (~1,030 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/blog/gemini-4-neon-more-horses-in-the-race.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-18

By Vlad Cealicu*, 18 November 2026. 873 words, about 4 min to read. * With the editing help of Claude Opus or Claude Fable.

- 61: Neon's score on the Artificial Analysis Intelligence Index at xhigh, third overall
- 4: Gemini 4 models Google has released in seven weeks
- $6.44: Neon's cost per Intelligence Index task, about 17% less than Gemini 4 Carbon at its standard price
- 110: output tokens a second, against about 57 for Carbon

Gemini 4 Neon. Sources [1] and [2].

Google released Gemini 4 Neon on 18 November 2026, its fourth Gemini 4 model in seven weeks [[2]](#ref-2). At its highest effort it scores 61 on the Artificial Analysis Intelligence Index, a point ahead of Gemini 4 Carbon and third overall, behind Claude Fable 5.2 and GPT-6.2 Astra [[2]](#ref-2). OpenAI released GPT-6.2 Astra the day before [[3]](#ref-3). Three labs, within a few points of each other, in the same week.

At the end of September Artificial Analysis wrote that Google was "back as one of the top three labs in intelligence achieved", after Gemini 4 Argon, its first model above the Flash class in more than seven months [[4]](#ref-4). Seven weeks later it has four models out and one of them is third on the index. It's cool to have more horses in the race. Nobody at the front gets to coast, prices come down, and anyone building agents gets a real choice of which model to use and which company to depend on.

## What Google shipped

Neon sits between Hydrogen, the fast everyday model, and Carbon, the one Google recommends for its hardest work [[5]](#ref-5). It reads text, images, video, audio and PDFs, writes text, holds a million tokens of context and has four effort levels, from low to xhigh [[2]](#ref-2). It's on the paid Gemini API in Google AI Studio, in Gemini for business, in Antigravity for Pro and Ultra subscribers and in the Gemini app [[1]](#ref-1), where it replaces Argon and Gemini 3.1 Pro [[5]](#ref-5).

The demo worth a look is a silly one. In Antigravity a swarm of Neon agents turned a single image into a 100,000-cell Google Sheet, with no embedded images and no Apps Script, using more than 500,000 live formulas to run edge detection, map depth and relight the scene across six modes [[1]](#ref-1).

## What it costs

Neon is $3 per million input tokens and $15 per million output [[2]](#ref-2). That's 25% below Carbon's standard $4 and $20, and 50% above Carbon's discounted $2 and $10, so until that discount ends on 30 November Carbon is the cheaper one by the token [[2]](#ref-2).

At Carbon's standard price Neon wins per task anyway. Artificial Analysis puts one of its Intelligence Index tasks at $6.44 on Neon, about 17% less than Carbon at its standard price, and about 21 minutes against 36 [[2]](#ref-2). Google's own figure is 32% less per task than Carbon at the default high effort [[1]](#ref-1), and on ExploitGym it reports frontier-level results with half of Carbon's tokens [[1]](#ref-1). Artificial Analysis found the opposite on its own index, about 15% more output tokens per task than Carbon, so there the saving comes from the lower price and from speed, about 110 output tokens a second against 57 [[2]](#ref-2). The price you pay for xhigh is a wait (56.51 seconds to the first answer token in Artificial Analysis's measurement) [[6]](#ref-6).

## Where it wins and where it doesn't

The number I'd show anyone picking a coding agent is Artificial Analysis's Coding Agent Index. Neon in Antigravity CLI scores 59, level with Claude Code running Claude Fable 5.2 (58) or Claude Opus 5.5 (58), for a fraction of the cost per task [[2]](#ref-2).

Google's own table puts Neon against Claude Opus 5.5 and GPT-6.2 Sol, and Neon leads on 16 of its 19 rows, mostly by a few points [[1]](#ref-1). The big gap is long context, 95.2% on GraphWalks from 256K to 1M tokens against 66.8% for Opus and 80.1% for Sol [[1]](#ref-1). On SWE-sweep it gets 39.0% in Google's own harness and 12.8% on the mini-SWE harness, where Opus 5.5 gets 3.6% [[1]](#ref-1).

The same table shows where it loses. Opus 5.5 is ahead on Terminal-Bench 4.1 (65.8% against 65.2%) and on FrontierCode 2.0 (47.3% against 46.2%), and GPT-6.2 Sol is ahead on Chartography (75.1% against 73.5%) [[1]](#ref-1). The two models above Neon on Artificial Analysis's index, Fable 5.2 and GPT-6.2 Astra, aren't in Google's table at all.

## What we do with it

Gemini 4 [Carbon](/tools/gemini-4-carbon) and [Hydrogen](/tools/gemini-4-hydrogen) are in our directory, graded against the same checklist as every other model API, and Neon isn't yet. It'll get a listing once we've read its model card, its pricing and its terms the way we read everyone's, and the vendor's own benchmarks go in marked as the vendor's, next to the independent ones.

What's still missing for me is open weights. None of the three horses at the front publishes its weights, so the race I want next is one where somebody running a box in their own room can stay within sight of the leaders.

Three closed labs trading points is a lot healthier than one running away with it. A year of Google, Anthropic and OpenAI pushing each other the way they did in the week of 16 November 2026 is a year of cheaper, faster and better agents for the rest of us, and I'd take a fourth horse too.

---

More in [Opinion](https://www.anchorterminal.com/blog/opinion/index.md)

- Older: [Did people sell bitcoin to buy Anthropic?](https://www.anchorterminal.com/blog/did-people-sell-bitcoin-to-buy-anthropic.md)
- Newer: [Closing the doors on MCP and calling it security](https://www.anchorterminal.com/blog/closing-the-doors-on-mcp.md)
- All posts: https://www.anchorterminal.com/blog/index.md

Anchor Terminal grades the tools AI agents run on, with a source for every score. [The top tools](https://www.anchorterminal.com/top/index.md) · [Categories](https://www.anchorterminal.com/categories/index.md) · [Compare two tools](https://www.anchorterminal.com/compare/index.md) · [Prices](https://www.anchorterminal.com/prices/index.md) · [Sunsets](https://www.anchorterminal.com/sunsets/index.md) · [API, MCP server and CLI](https://www.anchorterminal.com/agents/index.md)

## References

Sources read on 18 November 2026.

[1] Google, Introducing Gemini 4 Neon, 18 November 2026, with its benchmark table. Source for where Neon is available, the spreadsheet demo, the 32% lower cost per task than Carbon at high, ExploitGym with half of Carbon's tokens, and the table's figures for Neon, Claude Opus 5.5 and GPT-6.2 Sol, Neon leading on 16 of its 19 rows, GraphWalks, SWE-sweep, Terminal-Bench 4.1, FrontierCode 2.0 and Chartography. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/

[2] Artificial Analysis, Google continues to accelerate its release cadence with Gemini 4 Neon, 18 November 2026. Source for the fourth model in seven weeks, 61 at xhigh and third behind GPT-6.2 Astra and Claude Fable 5.2, a point ahead of Carbon, $6.44 and about 21 minutes per task against $7.76 and about 36 for Carbon, about 15% more output tokens, about 110 against 57 output tokens a second, the Coding Agent Index scores and costs per task, the inputs and context window, the four effort levels, and the prices against Carbon's. https://artificialanalysis.ai/articles/gemini-4-neon-intelligence-gain

[3] Artificial Analysis, OpenAI has the #2 model in the Intelligence Index with the release of GPT-6.2 Astra and GPT-6.2 Sol, 17 November 2026. Source for GPT-6.2 Astra's release the day before Neon. https://artificialanalysis.ai/articles/openai-has-the-2-model-in-the-intelligence-index-with-the-release-of-gpt-6-2-astra-and-gpt-6-2-sol

[4] Artificial Analysis, Gemini 4 Argon, Google is back as one of the top three labs in intelligence achieved, 30 September 2026. Source for the quotation and Argon being Google's first proprietary model above the Flash class in over seven months. https://artificialanalysis.ai/articles/gemini-4-argon-google-top-three-labs

[5] 9to5Google, Google launches Gemini 4 Neon with Carbon quality at Flash speeds, live for AI Pro, 18 November 2026. Source for Neon sitting between Hydrogen and Carbon and replacing Argon and Gemini 3.1 Pro in the Gemini app, as 9to5Google reports. https://9to5google.com/2026/11/18/gemini-4-neon-announcement/

[6] Artificial Analysis, Gemini 4 Neon, intelligence, performance and price analysis. Source for 56.51 seconds to the first answer token at xhigh. https://artificialanalysis.ai/models/gemini-4-neon

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
