Gemini 4 Neon Explained: Google’s New Mid-Tier Model, Price, Benchmarks and Who Should Use It

Sami Ullah Khan

November 20, 2026

Gemini 4 Neon Explained: Google's New Mid-Tier Model, Price, Benchmarks and Who Should Use It

Google released Gemini 4 Neon on 18 November 2026, the fourth model in its Gemini 4 family and arguably the one most businesses will end up using. Announced by Koray Kavukcuoglu, Google DeepMind’s chief AI architect, Neon is built for complex coding, analytical reasoning and production enterprise workflows, with what Google calls “frontier-level intelligence with fast end-to-end task completion.” It sits between the heavyweight Gemini 4 Carbon and the fast, inexpensive Gemini 4 Hydrogen, and Google’s pitch is simple: near-frontier results at a much lower cost per task. This guide covers what Neon is, how much it costs, how it performs, how it compares with Google’s other models and with rivals from OpenAI and Anthropic, how to access it and who should use it. Unless stated otherwise, details and benchmark figures come from Google’s announcement and are self-reported.

Gemini 4 Neon at a glance

DetailGemini 4 Neon
Released18 November 2026
Model IDgemini-4-neon
Input price$3 per 1M tokens
Cached input$0.15 per 1M tokens (a 95% discount)
Output price$15 per 1M tokens
Context window1 million tokens
Thinking levelsFour, to trade intelligence against token spend
Best forComplex coding, analysis, finance, legal and enterprise workflows

Where Neon fits in the Gemini 4 family

Google now has four Gemini 4 models, released over seven weeks. Google describes Carbon as the model for the hardest long-horizon projects, Hydrogen as the model for fast everyday execution and Neon as the balance of depth and efficiency.

ModelReleasedPositioningAPI price per 1M tokens (input / output)
Gemini 4 Argon13 Oct 2026General software development and knowledge work$2 / $10 introductory, then $4 / $20
Gemini 4 Carbon10 Nov 2026Hardest long-horizon coding, research and security tasks$2 / $10 introductory, then $4 / $20
Gemini 4 Hydrogen12 Nov 2026Fast, low-cost everyday coding and agents$1.50 / $7.50
Gemini 4 Neon18 Nov 2026Balance of depth and efficiency for complex and enterprise work$3 / $15

The introductory prices for Argon and Carbon run until 30 November 2026, as we covered in our Argon vs Carbon vs Hydrogen comparison. At first glance Neon’s $3 input price looks higher than Carbon’s introductory $2, but two factors favour Neon. Carbon’s price rises to $4 / $20 from December, and Carbon uses more tokens per task because it thinks longer. Google says that at its default “high” thinking level, Neon costs 32% less per task than Carbon.

How Neon compares on price with rival models

ModelInput per 1M tokensOutput per 1M tokens
Gemini 4 Neon$3.00$15.00
Claude Fable 5.2$8.00$40.00
GPT-6.2 Astra$8.00$40.00
Claude Opus 5.5$4.00$20.00
GPT-6.2 Sol$2.00$10.00
Claude Sonnet 5.5$2.00$10.00

Rival prices are from our coverage of Claude Fable 5.2 and GPT-6.2 Astra and Perplexity’s published Orchestrator API rates. Neon costs less than half as much as the two flagships per token. Google also claims that at its high setting Neon delivers roughly 50% to 80% lower cost per task than top competing frontier models, and at its highest xhigh setting at least 36% lower cost per task than competitors at their highest settings.

Benchmarks: general intelligence and agents

  • Artificial Analysis Intelligence Index: 61 at the xhigh thinking level, ranked third overall behind larger, slower models. For context, Claude Fable 5.2 scores 65 and GPT-6.2 Astra 62.
  • Vals Index: second at 70.9%, on the Pareto frontier of cost and performance.
  • Agent Arena Index: second globally with a score of 1336 across more than 4,200 sessions, and second on steerability, meaning how well it adapts when a user corrects it mid-task.
  • Agents’ Last Exam v1.1: 75.8%, on par with leading models.
  • Zapier AutomationBench: first at 56.3%.
  • Andon Labs Vending-Bench 2: first, with a simulated business result of $16,843.

Benchmarks: coding

BenchmarkGemini 4 Neon
Artificial Analysis Coding Agent Index58.6 (first)
Terminal-Bench 4.165.2%
FrontierCode 2.0 (default high setting)46.2%
SWE-sweep (in Google Antigravity)39.0%, described as state of the art
SWE-sweep (mini-SWE harness)12.8%

The FrontierCode 2.0 score of 46.2% compares with 49.1% that Anthropic reports for Claude Fable 5.2 at its highest effort, so Neon comes close to the leading flagship on this measure at a fraction of the price. The large gap between Neon’s two SWE-sweep scores shows how much results depend on the harness: in Google’s own Antigravity environment, with its tools and orchestration, the model performs far better than in a minimal setup.

Benchmarks: science, finance and legal work

  • Science: 70.5% on Terminal-Bench Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond, all described by Google as best in class.
  • Finance and spreadsheets: first on most finance, accounting, spreadsheet modelling and constraint-following benchmarks from Surge AI, and 73.5% on Chartography.
  • Legal and tax: first on Vals’ Legal Research Bench and Tax Agent Bench.

Google describes Neon as turning “inputs into structured deliverables”, which matches these results. Tasks such as building a financial model from raw data, researching a legal question or producing a structured report are where Neon’s benchmark lead is clearest.

Benchmarks: cybersecurity

  • SEC-bench Pro: 90.5%
  • AutoBaxBench (hard tier): 76%, state of the art
  • WebExploitBench v1.2: 51.26%, state of the art at half the cost of competing large models
  • CWE-bench v2: 70.7%, 3.4% above Carbon

Because strong security capability can be misused, Google is rolling out a cyber-focused variant of Neon only to vetted defenders through its Fairwind Program; teams request access through their Google Cloud account team.

Safety and reliability results

Google reports a set of behaviour and safety results for Neon:

  • 0.0% cheating on Terminal-Bench 4.0, matching Carbon
  • Shortcuts on 2.45% of tasks in a coding-cheating evaluation, the lowest of any Gemini 4 model
  • A 0.09% misaligned-action rate on Google’s agentic safety suite
  • A 99.2% pass rate on its agent honesty evaluation
  • A 0.0% attack success rate on Gray Swan’s indirect prompt-injection benchmark at the highest thinking level
  • Refusals on 1.98% of Terminal-Bench 4.0 tasks and none on Terminal-Bench Science
  • The lowest fabrication rate yet in Google’s internal Antigravity testing

The prompt-injection result is particularly relevant for agents. Indirect prompt injection, where malicious instructions are hidden in web pages or documents an agent reads, is one of the main ways AI agents can be hijacked. A model that resists it is safer to connect to email, browsers and business systems.

Thinking levels: how to choose

Neon offers four thinking levels that control how much reasoning the model does before answering. More thinking generally means better results on hard problems but more tokens and higher cost. Google’s benchmark comparisons use two of them: the default “high” level, where it says Neon sits on the cost-performance frontier, and the maximum “xhigh” level used for its Artificial Analysis score. A practical approach:

  • Use lower levels for simple extraction, formatting and short answers.
  • Use the default high level for most coding, analysis and agent work.
  • Reserve xhigh for the hardest problems where accuracy matters more than cost.

Where and how to use Gemini 4 Neon

SurfaceWho can use it
Gemini API (Google AI Studio, paid tier)Developers, with the model ID gemini-4-neon
Android StudioDevelopers
Gemini for businessBusiness customers via the API and model picker
Google AntigravityPaid Google AI Pro and Ultra subscribers
Gemini appGoogle AI Pro and Ultra subscribers

To try it as a developer, open Google AI Studio, choose Gemini 4 Neon from the model list and test prompts before moving to the API. For coding work, Antigravity, Google’s agentic development environment, is where Neon posts its strongest software-engineering results. In the Gemini app, Pro and Ultra subscribers can select Neon from the model picker.

Practical use cases

  • Codebase work: refactoring, debugging and implementing features across large repositories, using the 1 million-token context to hold more code at once.
  • Financial modelling: turning raw figures into structured spreadsheets and models, an area where Google reports leading scores.
  • Legal and tax research: finding and summarising relevant rules and precedents, with human review of every conclusion.
  • Business automation: multi-step workflows across apps, reflected in its first-place Zapier AutomationBench score.
  • Data analysis and reporting: converting datasets and documents into charts, summaries and structured deliverables.

Gemini 4 Neon vs Claude Fable 5.2 vs GPT-6.2 Astra

On the Artificial Analysis index, Neon (61) trails Claude Fable 5.2 (65) and GPT-6.2 Astra (62), and on FrontierCode 2.0 it trails Fable 5.2 slightly. But it costs $3 / $15 per million tokens against $8 / $40 for both flagships. For most production workloads, the question is not which model is best in absolute terms but which delivers acceptable quality at the lowest cost per completed task. On that measure, Neon is a strong contender, especially for finance, legal and spreadsheet-heavy work where Google reports first-place scores. The flagships remain the better choice for the hardest open-ended research and engineering problems.

Limitations and caveats

  • Self-reported results: most figures come from Google’s announcement. Independent evaluations will take time to appear.
  • Harness sensitivity: results vary widely with the tools around the model, as the two SWE-sweep scores show.
  • No free consumer access: in the Gemini app and Antigravity, Neon requires a paid Pro or Ultra subscription.
  • Restricted cyber variant: the security-focused version is available only to vetted organisations.

Who should use Gemini 4 Neon?

Neon is a strong default for teams that want near-frontier coding and analysis without paying flagship prices, and for organisations already using Google Cloud or Google Workspace. Choose Carbon for the very hardest long-horizon problems and Hydrogen when speed and cost matter most. If you are comparing vendors, run the same real tasks through Neon, Claude and GPT models and compare cost per finished task, not just price per token.

Why Google released four models in seven weeks

Google’s Gemini 4 rollout has been unusually fast. Argon arrived on 13 October, Carbon on 10 November, Hydrogen on 12 November and Neon on 18 November. Rather than one flagship with a cheaper “mini” version, Google has built a range in which each model targets a different balance of capability, speed and cost. Neon fills the gap that many businesses care about most: work that is too complex for a fast, cheap model but too frequent and high-volume to run on the most expensive one.

The timing also reflects competition. OpenAI released GPT-6.2 Astra and Sol on 17 November, and Anthropic released Claude Fable 5.2 on 9 November. Neon’s launch, one day after GPT-6.2, lets Google offer a lower-priced alternative that it says is close to those flagships on many tasks.

What 1 million tokens of context means in practice, and what it costs

A 1 million-token context window lets Neon read very large inputs in one go: an entire mid-sized codebase, hundreds of pages of contracts or a year of financial reports. That matters for tasks where the model needs to see how many parts fit together, such as tracing a bug across files or checking consistency across a long document.

Context caching makes repeated work on the same material much cheaper. Here is a simple worked example using Neon’s published prices:

ScenarioCalculationCost
Send a 200,000-token codebase as fresh input200,000 × $3 / 1M$0.60
Send the same codebase from cache200,000 × $0.15 / 1M$0.03
Generate a 5,000-token answer5,000 × $15 / 1M$0.075

If you ask 20 questions about the same cached codebase, input costs fall from about $12 to about $0.60, plus output. For teams that repeatedly query the same documents or repositories, caching can matter more to the final bill than the headline price per token. Note that thinking tokens, the model’s internal reasoning, are also billed, so higher thinking levels increase output costs.

Tips for getting good results from Neon

  • Be specific about the deliverable. Neon is tuned to produce structured outputs, so describe the format you want, such as a table, a spreadsheet layout or a report outline.
  • Give it the whole context. Use the large context window to include all relevant files or documents rather than excerpts, and cache material you will reuse.
  • Match the thinking level to the task. Start at the default and raise it only when results fall short.
  • Correct it mid-task. Neon ranks second on Agent Arena for steerability, so redirecting it during a long task is effective.
  • Verify important outputs. Even with low fabrication rates, check figures, legal conclusions and code before relying on them.

Key takeaways

  • Gemini 4 Neon is Google’s balanced model for complex coding, analysis and enterprise work, priced at $3 / $15 per million tokens.
  • It ranks third on the Artificial Analysis Intelligence Index and first on several coding, finance, legal and automation benchmarks, according to Google.
  • Google claims 32% lower cost per task than Carbon and 50% to 80% lower than competing frontier models at its default setting.
  • It is available through the Gemini API and to Google AI Pro and Ultra subscribers, but not to free users.
  • For the hardest problems, flagships such as Claude Fable 5.2 and GPT-6.2 Astra still lead; for high-volume professional work, Neon offers strong value.

What Neon means for businesses choosing an AI vendor

For companies deciding which AI models to build on, Neon reinforces a trend that has been building all year: the gap between the most capable model and a good-enough model is narrowing, while the price gap remains wide. A model that scores within a few points of the leaders at less than half the price changes the maths for any workload that runs thousands or millions of times a day.

That has three practical consequences. First, most organisations should use more than one model, routing routine work to cheaper models and escalating only the hardest tasks to flagships. Second, evaluation matters more than headlines: the best model for a legal research team may differ from the best model for a coding team, and only testing on real tasks reveals which. Third, switching costs are falling. Tools such as Google’s Antigravity Orchestrator, which mixes Carbon and Hydrogen in one session, and multi-provider APIs make it easier to change models as prices and capabilities shift.

For Google, Neon is a bid to win exactly this kind of high-volume enterprise work, where reliability, structured outputs and cost per task decide contracts more than leaderboard rankings do.

FAQ

How much does Gemini 4 Neon cost?

$3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens.

Is Gemini 4 Neon free?

No. In the Gemini app and Antigravity it requires a Google AI Pro or Ultra subscription; developers pay per token through the Gemini API.

What is the Gemini 4 Neon context window?

1 million tokens.

Is Neon better than Gemini 4 Carbon?

Carbon remains Google’s choice for the hardest long-horizon tasks, but Google says Neon costs 32% less per task at its default setting and scores higher on some benchmarks, such as CWE-bench v2.

What is the Gemini 4 Neon model ID?

gemini-4-neon.

Sources

Stay Ahead of AI

Get the latest AI news delivered to your inbox.

We don’t spam! Read our privacy policy for more info.

Published

Spotted an error? Read our corrections policy or email editorial@perplexityaimagazine.com.