What Gemini 4 Neon is, and when it came out
Gemini 4 Neon went GA on November 18, 2026, announced by Koray Kavukcuoglu of Google DeepMind. Google pitches it for "fast end-to-end task completion across complex software engineering, knowledge work, and science workflows". It's the fourth Gemini 4 model in about a month: Argon went GA on October 13, Carbon on November 10 and Hydrogen on November 12.
- Place in the family: Carbon for the hardest long-horizon projects, Hydrogen for fast everyday execution, and Neon for complex work you run often.
- Planner for Hydrogen: in Google Antigravity and through the Interactions API, Neon can plan and hand work to parallel Hydrogen subagents.
- Upgrade from Argon: Google's Cloud developer guide names Neon "the target upgrade" from Argon for general coding and enterprise reasoning, with migration steps from Gemini 3.1 Pro, Argon and 3.8 Flash.
- Gemini app: 9to5Google reports Neon takes Argon's place in the model picker, as "Deeper analysis". Argon is still listed in the Gemini API with prices, and we found no deprecation notice.
Gemini 4 Neon specs
| Gemini 4 Neon | |
|---|---|
| Released | November 18, 2026 (GA), by Google |
| Model ID | gemini-4-neon |
| Context window | 1,048,576 tokens. Google Cloud describes one limit shared by input and output. |
| Max output | 65,536 tokens per segment (AI Studio lists "64K / 1M"). On Google Cloud, long decoding (v1beta1 only) can continue up to the context limit minus input. |
| Knowledge cutoff | March 2026 |
| Input | Text, image, audio, video and PDF. Google Cloud adds Office and OpenDocument files (.docx, .pptx, .xlsx, .odt). |
| Output | Text only |
| Thinking levels | LOW, MEDIUM, HIGH (default), XHIGH |
| Supported (Google Cloud) | Thinking, system instructions, structured output, implicit caching (prefixes of 4,096+ tokens), grounding, code execution, function calling, URL context, computer use, agentic video understanding |
| Not supported (Google Cloud) | Explicit caching, tuning, Live API, RAG Engine |
| API price per 1M tokens | $3 input, $0.15 cached input, $15 output. No free tier. |
Gemini 4 Neon price vs Hydrogen, Carbon, Argon and Opus 5.5
Gemini API Standard prices per million tokens, from Google's pricing page on November 19, 2026. Output includes thinking tokens. Neon has no long-context tier and no dated promo. Carbon and Argon cost $2 / $10 through November 30 and $4 / $20 from December 1. So Neon costs more than both until November 30, and 25% less than both from December 1. Claude Opus 5.5 prices are Anthropic's.
Neon's Batch and Flex price is $1.50 input, $0.075 cached and $7.50 output. Priority is $5.40, $0.27 and $27. Google Search and Maps grounding each include 5,000 free requests a month (shared across Gemini 3 and later), then $14 per 1,000. Vertex AI per-token prices for Neon weren't readable when we checked, so this page gives Gemini API prices only.
| Model | Input | Cached input | Output | Free tier (API) |
|---|---|---|---|---|
| Gemini 4 Neon | $3 | $0.15 | $15 | No |
| Gemini 4 Hydrogen | $1.50 | - | $7.50 | Yes |
| Gemini 4 Carbon | $2 until Nov 30, then $4 | - | $10 until Nov 30, then $20 | No |
| Gemini 4 Argon | $2 until Nov 30, then $4 | - | $10 until Nov 30, then $20 | No |
| Claude Opus 5.5 | $4 | $0.20 | $20 | - |
Gemini 4 Neon benchmarks
Most of these are Google's own numbers from its launch post. Artificial Analysis rows are its independent runs. One row comes from Google's benchmark chart, which is an image, as reported by TestingCatalog. Google's model card is a PDF we couldn't read, so nothing here comes from it.
| Benchmark | Gemini 4 Neon | Reported by |
|---|---|---|
| Artificial Analysis Intelligence Index | 61 (xhigh), third at launch | Artificial Analysis |
| AA Coding Agent Index v2.1 | 59 (xhigh, in Antigravity CLI) | Artificial Analysis (Google quotes 58.6) |
| Vals Index | 70.9%, #2 | |
| Agent Arena Index (arena.ai) | 1336, #2 at debut (4,200+ sessions) | |
| Terminal-Bench 4.1 | 65.2% | |
| Terminal-Bench Science | 70.5% | |
| FrontierCode 2.0 (Cognition) | 46.2% at default high | |
| SWE-sweep | 12.8% (mini-SWE harness), 39.0% in Antigravity | |
| CursorBench 4.1 | 65.4% | Google's chart, via TestingCatalog |
| HLE-Diamond | 71.3% | |
| Agents' Last Exam v1.1 | 75.8% | |
| RiemannBench | 84.0% | |
| Zapier AutomationBench | 56.3%, #1 overall | |
| Andon Labs Vending-Bench 2 | $16,843, #1 overall | |
| Chartography | 73.5% | |
| AutoBaxBench, hard tier | 76% | |
| SEC-bench Pro | 90.5% | |
| CWE-bench v2 | 70.7% (+3.4% over Carbon) | |
| WebExploitBench v1.2 | 51.26% |
What else Google reports
- Vals: Google says Neon is #1 on Legal Research Bench and Tax Agent Bench, and #2 for steerability on the Agent Arena.
- Surge AI: Google says Neon is #1 on most of Surge AI's finance, accounting, spreadsheet-modeling and constraint-following benchmarks. It gave no numbers.
- Cost per task: Google says Neon costs 32% less per task than Carbon at the default high setting, and at xhigh costs 36%+ less per task than "competing models" at their highest settings.
- Safety, per Google: refusals on 1.98% of Terminal-Bench 4.0 tasks, a 0.0% cheating rate on Terminal-Bench 4.0, shortcuts on 2.45% of a coding cheating eval (lowest in Gemini 4), misaligned actions at 0.09%, and 99.2% on an honesty eval. Against Carbon, Google reports 43–56% less under-service and 15–35% less overreach.
- Neon for cyber is rolling out to eligible defenders in Google's Fairwind Program over the coming days.
Gemini 4 Neon on Artificial Analysis
Artificial Analysis (AA) ran Neon independently and published an article on launch day. These are its numbers.
- Intelligence Index: 61 at xhigh, third at launch, behind GPT-6.2 Astra (max) and Claude Fable 5.2 (max), and 1 point ahead of Carbon (high, 60). By thinking level: xhigh 61, high 60, medium 55, low 44. See GPT-6.2 and Claude Fable 5.2.
- Coding Agent Index (v2.1: FrontierCode-AA 2.0, Terminal-Bench 4.1, SWE-Atlas-QnA-AA): 59 at xhigh in Antigravity CLI, 3 points ahead of Carbon (high, 55). Claude Code with Fable 5.2 (xhigh) and Claude Opus 5.5 (max) both scored 58. Google calls Neon #1 on this index; AA's article states no rank.
- Cost per task on the Intelligence Index (xhigh): $6.44, about 17% below Carbon's $7.76 at non-discounted prices.
- Cost per task on the Coding Agent Index: Neon $6.86, Opus 5.5 (max) $12.61, Fable 5.2 (xhigh) $15.47.
- Speed: about 21 minutes per task against about 36 for Carbon, and about 110 output tokens per second against about 57. Neon uses about 15% more output tokens than Carbon.
- Other AA evals: GDP.pdf 33.4% at high and 28.0% at xhigh (Carbon high: 28.4%). AA-Briefcase 1700 Elo at xhigh, and #1 on Presentation Quality at 1771. MKA Climb WSR 56% (Fable 5.2 max: 57%).
Gemini 4 Neon on the Models Engineering Index (third party)
The Models Engineering Index (Mei) is a third-party leaderboard published on GitHub (git-aarvin/mei). It combines coding-benchmark results attributed to Cognition, Cursor, the Laude Institute and Vals AI into one score, scaled so the average model scores 50. As of Nov 20, 2026, Neon is #1.
- Index: Gemini 4 Neon 68.5, Claude Fable 5.2 66.0, GPT-6.2 Astra 62.8, Claude Opus 5.5 62.0, Claude Sonnet 5.5 58.6, Gemini 4 Hydrogen 57.5, GPT-6.2 Sol 56.4.
- Cost relative to a typical model: Neon 1.7×, Opus 5.5 3.6×, Fable 5.2 6.5×. Tokens used: Neon 0.82×, Opus 5.5 1.9×, Fable 5.2 2.1×.
- Mei's own caveats: models are scored on different subsets of benchmarks, there are no error bars, so small gaps mean little, and the results are not independently verified. Carbon isn't on the board.
Gemini 4 Neon vs Carbon vs Hydrogen vs Argon
Prices are Gemini API Standard rates per million tokens. AA scores are the highest Intelligence Index figure Artificial Analysis shows for each model. Google's own figure for Hydrogen is 51 at high and 54 at xhigh.
| Hydrogen | Neon | Argon | Carbon | |
|---|---|---|---|---|
| Google's role for it | Fast, everyday execution | Depth plus efficiency, complex work you run often | Efficient frontier model; Neon is its upgrade target | Hardest long-horizon projects |
| GA (Gemini API) | Nov 12, 2026 | Nov 18, 2026 | Oct 13, 2026 | Nov 10, 2026 |
| API price in / out | $1.50 / $7.50 | $3 / $15 | $2 / $10 to Nov 30, then $4 / $20 | $2 / $10 to Nov 30, then $4 / $20 |
| Free tier (API) | Yes | No | No | No |
| Thinking levels | low, medium, high, xhigh | low, medium, high (default), xhigh | up to high (default medium) | - |
| AA Intelligence Index | 54 (xhigh) | 61 (xhigh) | 50 | 60 (high) |
| AA Coding Agent Index | - | 59 (xhigh) | - | 55 (high) |
| Gemini app | Everyday help | Deeper analysis (AI Pro, Ultra) | Replaced by Neon in the picker (9to5Google) | Code & complex work (Ultra, per Google's Logan Kilpatrick via VentureBeat) |
Which Gemini 4 model to pick
- Neon vs Carbon: on Artificial Analysis, Neon scores 1 point higher on the Intelligence Index and 3 points higher on the Coding Agent Index, finishes tasks faster and costs about 17% less per task. Per token it costs more than Carbon until November 30 and 25% less from December 1. Google still keeps Carbon for the hardest long-horizon projects.
- Neon vs Hydrogen: Neon costs twice as much per token and scores 7 points higher on AA's Intelligence Index (61 vs 54). Hydrogen has a free API tier and is the fast worker; Neon can plan and hand work to Hydrogen subagents.
- Neon vs Argon: Google names Neon the upgrade target. Neon defaults to HIGH thinking and supports XHIGH, where Argon defaults to MEDIUM and tops out at HIGH. Neon's output segment is 65,536 tokens against 32,768 on Argon, and Neon takes Office and OpenDocument input. Argon's
europe-west2andasia-southeast1locations aren't supported for Neon on Google Cloud. - Rule of thumb: Hydrogen for fast, cheap, high-volume work; Neon for most complex coding and analysis; Carbon when a long project still stalls on Neon.
Gemini 4 Neon vs Claude Opus 5.5 for coding
Only benchmarks where Google and Anthropic each published their own number, plus Artificial Analysis's independent runs of both. The vendor figures are vendor-reported, with different harnesses, effort settings and safeguards, so they aren't head-to-head runs. Press reports quote other Opus 5.5 numbers from Google's benchmark chart, which don't all match Anthropic's own; we leave those out. For OpenAI's models against Claude, see GPT-6.2 vs Claude.
| Gemini 4 Neon | Claude Opus 5.5 | Note | |
|---|---|---|---|
| AA Coding Agent Index | 59, xhigh, Antigravity CLI (AA) | 58, max (AA) | Same referee. Cost per task: Neon $6.86, Opus $12.61. |
| FrontierCode 2.0 | 46.2%, default high (Google) | 45.5%, highest effort, safeguards on (Anthropic) | Vendor-reported |
| Terminal-Bench-Science | 70.5% (Google) | 58.7%, v0.1 (Anthropic) | Vendor-reported |
| Zapier AutomationBench | 56.3% (Google) | 42.5% (Anthropic) | Anthropic counts safeguard interventions as failures |
| Chartography | 73.5% (Google) | 64.4%, no tools (Anthropic) | Google doesn't say whether Neon used tools |
| API price in / out per 1M | $3 / $15 | $4 / $20 | Neon is 25% lower. Cached input: $0.15 vs $0.20. |
| Context | 1,048,576 tokens, shared by input and output | 1M input, 128K output | - |
What that means for coding
- On the one same-referee coding number, AA's Coding Agent Index, Neon (59) and Opus 5.5 (58) are a point apart, and Neon's runs cost about half as much per task.
- On the four vendor-reported benchmarks, Google's Neon figures are higher than Anthropic's own Opus 5.5 figures. The setups differ, so test both on your own tasks.
- The agent matters too. Opus 5.5 runs in Claude Code; Neon runs in Google Antigravity and through the Gemini API. See Claude Code vs Codex for how the two main coding agents compare.
Where you can use Gemini 4 Neon
- Gemini API and Google AI Studio: on the paid tier only. Google says all AI Pro and Ultra subscribers can try it in AI Studio.
- Google Antigravity: for paid Google AI Pro and Ultra subscribers.
- Gemini app: AI Pro and AI Ultra subscribers can pick it in the model dropdown. 9to5Google says it powers "Deeper analysis" and multi-step Tasks. Google's sources don't name Free or AI Plus.
- Android Studio and Gemini for business (through the API and the model dropdown).
- Google Cloud (Gemini Enterprise Agent Platform, the Vertex AI docs): GA at
global,usandeu, for verified enterprise customers in select regions. Billing is Standard PayGo, Priority PayGo, Provisioned Throughput or batch. Cache writes cost 1.25x the input price. - Gemini CLI: Google's launch post doesn't mention it, and Gemini CLI's release notes don't list Neon or any Gemini 4 model as of November 19.
How to use Gemini 4 Neon
Try it in Google AI Studio
Pick Gemini 4 Neon in AI Studio. It isn't on the free tier, so you need a paid API key or an AI Pro or Ultra subscription.
Call it from the Gemini API
Use the model ID
gemini-4-neon. Thinking defaults to HIGH. Use XHIGH for the hardest tasks, and MEDIUM or LOW when speed matters more.# Gemini API: model ID gemini-4-neon (paid tier only) curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-4-neon:generateContent" \ -H "x-goog-api-key: $GEMINI_API_KEY" \ -H "Content-Type: application/json" \ -d '{"contents": [{"parts": [{"text": "Find and fix the failing test in this diff"}]}]}' # thinking levels: LOW | MEDIUM | HIGH (default) | XHIGH # thinking_budget or MINIMAL returns HTTP 400; temperature, top_p and top_k are ignoredCode with it in Google Antigravity
On a paid AI Pro or Ultra plan, choose Neon in Antigravity. There it can plan and hand work to parallel Hydrogen subagents.
Moving from Argon? Check the differences
Default thinking goes from MEDIUM to HIGH.
thinking_budgetandMINIMALare rejected. On Google Cloud,europe-west2andasia-southeast1aren't supported for Neon.
Demos Google described
Google's launch post describes four builds made with Neon. None can be played: the post has no "try it" link, and we found no hosted build or repo.
The first Neon game you can play came a day later: on Nov 19 Google posted NEON SCRAMBLE, a 3D runner where you guide Google Maps' Pegman from Shibuya Scramble Crossing to the Google building in Tokyo. Google says Neon turned nearly 1M tokens of real OpenStreetMap data (654 roads, 70 rail lines and 280 building footprints) into the map. Play it or see it on vibegamer.io.
Early testers' Neon games and the first playable ones are tracked in Gemini 4 Neon games on vibegamer.io; for Google's world model, see Project Genie games.
For games specifically, vibegamer.io compares Neon with Claude Opus 5.5 and Fable 5.2 in Gemini 4 Neon vs Opus 5.5 vs Fable 5.2 for making games, with a playable game from each model.
- Ski simulator: a first-person WebGL2 freeride ski simulator written from scratch on real Mont Blanc / Chamonix terrain, with 10,200 lines of custom shaders, 120 Hz ski physics on six lines, slab avalanches, spatial Web Audio and no pre-made assets.
- Turing's study: Alan Turing's 1936 study at King's College, Cambridge, rebuilt in interactive 3D from archival photos with /swarm, with a working Universal Turing Machine on the desk.
- Lab bench: a photorealistic 3D lab bench built through Blender MCP in Antigravity, with Bunsen flames for Carbon and Hydrogen and gas discharge tubes for Argon and Neon.
- Spreadsheet swarm: one image turned into a 100,000-cell Google Sheet with 500,000+ live formulas for edge detection, depth mapping and relighting, without images or Apps Script.
A new model is half of it. The other half is a job.
Models keep getting better and cheaper, and the agents around them still wait for the next prompt. Crewmark turns coding agent sessions into AI employees with a role, a goal in numbers, working hours and a heartbeat, so they take the next issue from a board on their own. Claude Code and Codex employees run natively; other agents, such as OpenCode, DeepSeek Harness and Gemini, connect through ACP (acpx). An employee uses whatever model its agent runs.
Pick a role
From the AI employee roles, such as an AI software engineer.
Put the work on a board
Assign issues, make the regular ones repeat, and let the heartbeat wake the employee to take the next one.
See what it can do
Jobs with setup and rules for each agent are in Claude Code use cases and Codex use cases.
FAQ
What is Gemini 4 Neon?
A Gemini 4 model Google released on November 18, 2026, for complex software engineering, knowledge work and science. It sits between Carbon and Hydrogen, with a 1M-token context, text-only output and four thinking levels. The model ID is gemini-4-neon.
How much does Gemini 4 Neon cost?
In the Gemini API, $3 per million input tokens, $0.15 cached and $15 output, including thinking tokens. Batch and Flex are $1.50, $0.075 and $7.50. There is no free API tier. In the Gemini app it needs AI Pro or AI Ultra.
What are Gemini 4 Neon's benchmarks?
Google reports 65.2% on Terminal-Bench 4.1, 70.5% on Terminal-Bench Science, 46.2% on FrontierCode 2.0, 70.9% on the Vals Index and 75.8% on Agents' Last Exam v1.1. Artificial Analysis independently measured 61 on its Intelligence Index and 59 on its Coding Agent Index.
How did Gemini 4 Neon do on Artificial Analysis?
At launch it placed third on the Intelligence Index with 61 (xhigh), behind GPT-6.2 Astra and Claude Fable 5.2 and 1 point ahead of Carbon. On the Coding Agent Index it scored 59, against 58 for Claude Opus 5.5 and for Claude Code with Fable 5.2, at $6.86 per task.
Gemini 4 Neon vs Carbon: which is better?
On Artificial Analysis, Neon scores slightly higher than Carbon (61 vs 60 on the Intelligence Index, 59 vs 55 on the Coding Agent Index), runs faster and costs about 17% less per task. Google still positions Carbon for the hardest long-horizon projects. From December 1, Neon is also 25% cheaper per token.
Gemini 4 Neon vs Hydrogen or Argon?
Hydrogen is the fast, cheaper model at $1.50 / $7.50 with a free tier. Neon costs twice as much and scores higher on AA's Intelligence Index (61 vs 54). Google calls Neon the upgrade target for Argon: it adds XHIGH thinking, a larger output segment and Office file input. See Gemini 4 Hydrogen.
Is Gemini 4 Neon better than Claude Opus 5.5?
On Artificial Analysis's Coding Agent Index, Neon scored 59 and Opus 5.5 scored 58, with Neon at about half the cost per task. On FrontierCode 2.0, Terminal-Bench-Science, AutomationBench and Chartography, Google's own Neon figures are higher than Anthropic's own Opus figures, but these are vendor-reported with different setups. Neon is 25% cheaper per token.
Can I use Gemini 4 Neon in Gemini CLI?
Google hasn't said. The launch post doesn't mention Gemini CLI, and its release notes don't list Neon as of November 19, 2026. Neon is in AI Studio, the Gemini API, Antigravity, Android Studio, the Gemini app and Google Cloud.
Checked against Google's launch post, the Gemini API pricing, models and changelog pages, Google Cloud's Neon docs, Anthropic's Claude pages and Artificial Analysis on November 19, 2026.