SPIRITT logoSPIRITTGoogle DeepMindGemini 4 Neon

Gemini 4 Neon. Near-Carbon depth, built to finish the job.

Google's new middle Gemini 4 model sits between Hydrogen and Carbon. At its default high thinking level it scores within half a point of Gemini 4 Carbon on the Artificial Analysis Intelligence Index, leads Zapier's business-workflow benchmark and Cursor's coding benchmark, and costs $3 input and $15 output per million tokens. In SPIRITT, put it on complex code, dense documents and agents that run for hours.

Gemini 4 Neon across X

What people say after putting it to work: developers who made it their daily driver, a 3D build from a single image, a bug hunt on real repos, and the independent tests behind the buzz. Not all praise: one early tester says its one-shot demos look bad.

The Gemini 4 for hard work that has to finish

Neon brings Carbon-class reasoning to work you run every day. Independent evaluators confirm the business-workflow, coding and document results, and show where it costs more or trails.

What Google launched, and what it means in SPIRITT

Google released Gemini 4 Neon on November 18, 2026, generally available on the paid tier of the Gemini API, in Google AI Studio, Gemini for business, the Gemini app for Google AI Pro and Ultra subscribers, and Google Antigravity for paid Pro and Ultra subscribers. It is not on the free tier. Google places it between Gemini 4 Hydrogen and Gemini 4 Carbon for complex work that runs often: Carbon for the hardest architecture, debugging, science and security problems, Neon for complex software engineering, knowledge work and science with fast completion, and Hydrogen as the everyday default.

Google adapted Neon from Gemini 4 Carbon and built it for long-running agent work: repository-level debugging and multi-file refactoring in fewer steps, fewer redundant file reads, reliable context compaction in long sessions, and turning spreadsheets, PDFs, charts and recordings into structured deliverables. Google also positions it as an orchestrator that directs parallel Hydrogen workers.

Official limits: one 1,048,576-token window shared by input, thinking and output, up to 65,536 output tokens per response and up to about 1M with long decode, text, image, video, audio and PDF input, text output, and a March 2026 knowledge cutoff. Google offers four thinking levels: low, medium, high (the default) and xhigh.

On the Artificial Analysis Intelligence Index v5.0.1, Neon scores 59.6 at high, next to Gemini 4 Carbon (60.0), and 61.4 at xhigh, behind only Claude Fable 5.2 and GPT-6.2 Astra at their top settings. At high it leads every other model on Zapier's AutomationBench (54.3%) and Cursor's CursorBench 4.1 (63.7%), and posts the best GDP.pdf document score Artificial Analysis has measured. It still trails on terminal coding, and GPT-6.2 Astra and Sol reach similar intelligence faster and for less per task.

Google's API price is $3.00 per million input tokens and $15.00 per million output tokens across the whole 1M window, with cached input at $0.15, implicit caching only. Batch and Flex run at half price; Priority costs $5.40 and $27.00. There is no introductory discount. Hydrogen costs half as much; Carbon is $2 / $10 through November 30, then $4 / $20. Thinking tokens bill as output.

Gemini 4 Neon is available in SPIRITT Workspaces with low, medium and high thinking; Google's xhigh level is not offered in SPIRITT yet. Start at high for complex work, try medium for interactive turns, and keep consequential approvals with a person.

AA Index 59.6 at high#1 AutomationBench#1 CursorBench 4.11M shared context$3 / $15Available in SPIRITT

AA Intelligence Index v5.0.1

Claude Fable 5.2 (max)
64.6
GPT-6.2 Astra (max)
62.1
Gemini 4 Neon (xhigh)
61.4
Gemini 4 Carbon (high)
60
Gemini 4 Neon (high)
59.6
Claude Opus 5.5 (max)
59.1
GPT-6.2 Sol (max)
57.3
Claude Sonnet 5.5 (max)
54.5
Gemini 4 Hydrogen (xhigh)
53.9
Gemini 4 Argon (high)
50.4

Artificial Analysis, November 18, 2026 snapshot. Index points, not percentages. Neon is highlighted at high, the deepest level SPIRITT offers today; its xhigh row is Google's top level, not yet available in SPIRITT. Other models use their best published configuration unless labelled; Claude models run with Anthropic's default fallback.

Where Neon leads. Where the trade-offs are.

Business workflows, agents, coding, documents, reliability and long context, with independent evaluations first and Google's own table clearly marked. Cost, speed and the terminal-coding gap are shown too.

Independent business workflows
Winner

AutomationBench 1.0.6: Zapier

Gemini 4 Neon (high)
54.34%
Gemini 4 Carbon (high)
53.88%
GPT-6.2 Astra (max)
52.36%
Claude Fable 5.2 (max)
51.45%
Gemini 4 Argon (high)
51.29%
Gemini 4 Hydrogen (high)
51.14%

Zapier's strict pass rate on end-to-end tasks across 47 simulated business apps, at standard list pricing. Neon at high leads every other model for $0.88 per task, against $2.12 for Carbon and $4.64 for Fable 5.2; Neon at xhigh scores 56.32% for $1.69 and at medium 53.27% for $0.58. GPT-6.2 Astra leads the HR domain.

Independent agent sessions
Near frontier

Agent Arena Index

Claude Fable 5.2 (max)
1453
Gemini 4 Neon (high)
1336
Claude Opus 5.5 (max)
1335
Gemini 4 Carbon (high)
1303
Claude Sonnet 5.5 (max)
1297
GPT-6 Astra (max)
1280
Gemini 4 Hydrogen (high)
1271

Arena.ai's live Agent Arena Index, November 18, 2026, from 4.2k real agent sessions rated by Arena's community. Neon (High) ranks #2 overall and #2 on steerability, one point above Claude Opus 5.5 (Max), a statistical tie. Its $2.82 median cost per task is 69% below Fable 5.2 and 53% below Opus 5.5, but about twice Carbon's $1.34.

Independent coding sessions
Winner

CursorBench 4.1

Gemini 4 Neon (high)
63.7%
Gemini 4 Carbon (high)
61.5%
Gemini 4 Hydrogen (xhigh)
61.4%
Claude Fable 5.2 (max)
61.3%
Claude Opus 5.5 (max)
58.1%

Cursor's leaderboard of ambiguous, multi-file tasks from real Cursor sessions. Neon at high costs $6.54 per task against $8.52 for Carbon, $11.88 for Opus 5.5 and $48.11 for Fable 5.2; Neon at extra high scores 65.4% for $10.67. GPT-6.2 models are not on this board.

Independent long-document reasoning
Winner

GDP.pdf: all-pass rate

Gemini 4 Neon (high)
33.4%
GPT-6.2 Sol (max)
30%
Gemini 4 Hydrogen (high)
30%
GPT-6.2 Astra (max)
29.8%
Gemini 4 Carbon (high)
28.4%
Claude Opus 5.5 (max)
26.2%
Claude Fable 5.2 (max)
24.2%

Artificial Analysis's visual reasoning benchmark over long, complex documents. Neon at high posts the best score of any configuration Artificial Analysis lists; Neon at medium (32.8%) and low (31.4%) also beat every other model.

Fewer confident wrong answers
Near frontier

AA-Omniscience: hallucination rate

Gemini 4 Carbon (high)
11.4%
Gemini 4 Neon (high)
13.9%
Claude Sonnet 5.5 (max)
43.3%
Claude Opus 5.5 (max)
50.3%
GPT-6.2 Sol (max)
55.4%
GPT-6.2 Astra (max)
59.1%
Claude Fable 5.2 (max)
59.3%

Artificial Analysis's share of wrong answers among questions the model did not get right, lower is better, for the top-scoring configurations on the Index. Neon guesses a quarter to a third as often as its non-Gemini peers. Gemini 4 Argon (4.3%), Hydrogen (8.7%) and Carbon at medium (4.2%) guess even less. A benchmark measure, not a real-world error rate.

Cost at similar intelligence

AA cost per Index task, models scoring about 57 to 62

GPT-6.2 Sol (max)
$1.01
GPT-6.2 Astra (high)
$2.07
Gemini 4 Neon (high)
$3.74
Gemini 4 Carbon (high)
$4.90
Claude Opus 5.5 (max)
$10.04
Claude Fable 5.2 (xhigh)
$14.83

Artificial Analysis's weighted cost to run one Intelligence Index task at each provider's list price, lower is better. GPT-6.2 Sol and Astra reach similar intelligence for less. Neon at high costs about a quarter less than Carbon at Carbon's current introductory price. Not a SPIRITT price.

Speed at similar intelligence

Minutes per Index task, models scoring about 57 to 62

GPT-6.2 Astra (high)
5.9
GPT-6.2 Sol (max)
8.5
Gemini 4 Neon (high)
14.8
Claude Fable 5.2 (xhigh)
24.1
Claude Opus 5.5 (max)
24.6
Gemini 4 Carbon (high)
36.1

Artificial Analysis's weighted time per Intelligence Index task, lower is better. Neon at high finishes in less than half of Carbon's time and outputs about 100 tokens per second, but GPT-6.2 Astra and Sol are faster still. Neon at medium takes 8.3 minutes for a score of 54.9.

Agentic terminal work: a gap

Terminal-Bench 4.0: AA

Gemini 4 Carbon (high)
65.2%
GPT-6.2 Astra (max)
64.1%
Claude Sonnet 5.5 (max)
63.6%
Gemini 4 Neon (high)
61.6%
GPT-6.2 Sol (max)
60.1%
Claude Opus 5.5 (max)
59.6%
Claude Fable 5.2 (max)
59.6%

Artificial Analysis's terminal-agent run. Neon at high trails Carbon, GPT-6.2 Astra and Claude Sonnet 5.5; at xhigh it reaches 63.6%. Vals AI's own run at xhigh places it #7 of 69 at 61.62%.

Independent hard reasoning

HLE-Diamond: AA

Claude Fable 5.2 (max)
80.2%
GPT-6.2 Astra (max)
72.2%
Gemini 4 Carbon (high)
69.1%
Gemini 4 Neon (high)
68.4%
Claude Opus 5.5 (max)
67.9%
GPT-6.2 Sol (max)
64%

Artificial Analysis's run of the hardest Humanity's Last Exam subset. Neon at high sits in the frontier pack but behind Claude Fable 5.2 and GPT-6.2 Astra; at xhigh it reaches 71.3%, the figure Google cites as best in class against Opus 5.5 and GPT-6.2 Sol.

Long context: Google's table
Winner

GraphWalks BFS, 256K to 1M tokens

Gemini 4 Neon
95.2%
GPT-6.2 Sol
80.1%
Claude Opus 5.5
66.8%

Google DeepMind's model card, November 2026; vendor-reported and self-computed, not an independent run. Google runs Neon at its highest thinking setting (xhigh) unless noted, a level SPIRITT does not offer yet. Google compares Neon only with Claude Opus 5.5 and GPT-6.2 Sol.

Independent sources: Artificial Analysis Intelligence Index v5.0.1 and component evaluations (November 18, 2026 snapshot; Claude models with Anthropic's default fallback); Arena.ai Agent Arena (November 18); Zapier AutomationBench 1.0.6; Cursor's CursorBench 4.1; Vals AI. Vendor source: Google DeepMind's Gemini 4 Neon model card and evaluation methodology, November 2026, which run Neon at its highest thinking setting unless noted. Neon is shown at high unless a row says otherwise, because SPIRITT offers low, medium and high today. Effort levels, harnesses and scoring differ between boards; close scores do not establish a meaningful lead. Cost per task depends on each evaluator's price basis and is not a SPIRITT price.

How It Works

Choose Gemini 4 Neon in the picker, then hand it the complex work that needs depth and still has to finish.

01

Open a workspace

Open a workspace and land in a fully equipped cloud computer: browser, files, terminal, integrations, and memory. No local setup. No thin chat box pretending to be an agent.

Open a SPIRITT workspace
02

Choose Gemini 4 Neon

Open the SPIRITT model picker and select Gemini 4 Neon. Use it for multi-file coding and debugging, dense documents and spreadsheets, research and science work, and agents that run for hours, at low, medium or high thinking. Keep Gemini 4 Hydrogen for fast everyday turns and bring in Gemini 4 Carbon for the hardest problems.

Multicolored glass Gemini star circled by a softly glowing neon glass tube above a frosted glass plinth representing Gemini 4 Neon in the model picker
03

Build or automate

Tell it what to ship or what to run. Gemini 4 Neon can code, call tools, drive the browser, coordinate multi-step work, and keep going while you step away. The point is not another chat window. It is an agentic environment where Gemini 4 Neon actually does the job.

Build and automate with Gemini 4 Neon

Put near-Carbon reasoning on the work you run every day

Open a SPIRITT Workspace, choose Gemini 4 Neon, and give it the files, tools and checks for complex work that runs end to end.

Questions

Gemini 4 Neon FAQ

01What is Gemini 4 Neon?+
Google's middle Gemini 4 model, released on November 18, 2026. It sits between Gemini 4 Hydrogen and Gemini 4 Carbon and brings frontier-level reasoning to complex software engineering, knowledge work and science that runs often. Google adapted it from Gemini 4 Carbon.
02How is it different from Gemini 4 Carbon and Hydrogen?+
At high thinking it scores 59.6 on the Artificial Analysis Intelligence Index, next to Carbon (60.0) and well above Hydrogen (50.8), and finishes an Index task in less than half of Carbon's time. Google keeps Carbon for the hardest architecture, debugging, science and security problems and Hydrogen as the fast everyday default. Neon defaults to high thinking and adds xhigh, while Carbon defaults to medium and caps at high.
03How does it compare with Gemini 4 Argon?+
Google names Neon the upgrade target for general coding and enterprise reasoning on Argon. It scores 9 points higher on the Artificial Analysis Intelligence Index at high, leads Argon on Zapier's AutomationBench and Arena's Agent Arena, and from December 1 costs less per token than Argon's $4 / $20 standard price. Argon still guesses less often when it does not know an answer.
04What is it best for?+
Complex work you run again and again: repository debugging and multi-file refactors, business workflows across many apps, turning PDFs, spreadsheets, charts and recordings into finished reports, legal, finance and science analysis, and long-running agents, including coordinating faster Hydrogen workers.
05What are the main limitations?+
It trails Gemini 4 Carbon, GPT-6.2 Astra and Claude Sonnet 5.5 on terminal coding (61.6% on Terminal-Bench 4.0 at high), and Claude Fable 5.2 and GPT-6.2 Astra still score higher overall. GPT-6.2 Astra and Sol reach similar intelligence faster and for less per task, and on Arena's Agent Arena Neon costs about twice Carbon per task. It is paid tier only, costs twice Hydrogen per token, and one early tester's first run disappointed. Knowledge cutoff is March 2026.
06How much does Gemini 4 Neon cost?+
Google's API price is $3.00 per million input tokens and $15.00 per million output tokens, flat across the 1M window, with cached input at $0.15 per million and implicit caching only. Batch and Flex cost half; Priority costs $5.40 and $27.00. There is no introductory discount, and thinking tokens bill as output. These are Google API prices, not SPIRITT plan prices.
07What are its context and output limits?+
Input, thinking and output share one 1,048,576-token window. A standard response stops at 65,536 output tokens; Google's long decode continues a response up to about 1M tokens minus your input. It reads text, images, video, audio and PDFs and writes text.
08What is a good example of using it well?+
Give it a hard, repeatable job with a definition of done it can test, for example: reconcile this quarter's contracts, invoices and the finance model, flag every mismatch with its source, fix the spreadsheet formulas, and draft the follow-ups. Start at high thinking, move to medium where results hold, and check the final output yourself.
09Is Gemini 4 Neon available in SPIRITT Workspaces?+
Yes. Open a SPIRITT Workspace, choose Gemini 4 Neon in the model picker, and use it with browser, terminal, files, tools, integrations, memory and durable sessions. SPIRITT offers low, medium and high thinking; Google's xhigh level is not available in SPIRITT yet.
Buy from builders who use what they sellBuilt usingSPIRITT