← All AI resources

Google DeepMind · Gemini 4 Neon launch and evaluation guide

Gemini 4 Neon trades peak depth for frequent frontier work

Google placed Neon between Hydrogen and Carbon for complex, high-frequency work. Independent results support its capability claim, while showing that maximum reasoning can be slow and unusually verbose.[1][5]

Published by Learnetto team · Published November 19, 2026

Routine and fastHydrogenLower-cost execution for everyday coding, tools and multimodal work.
Complex and frequentNeonFour reasoning levels for repeated engineering and analytical workflows.
Hardest and longCarbonDeeper planning for the tasks where measured result quality earns the cost.
Route by accepted result, total billed tokens and elapsed time. Model family labels alone cannot choose the best operating point.

Google released Gemini 4 Neon on November 18, 2026 for complex coding, analytical work and agent workflows. It has four reasoning levels, a unified 1,048,576-token input-output window and standard API prices of $3 per million input tokens and $15 per million output tokens.[1][3][4] Artificial Analysis scored Neon at 61 on its Intelligence Index at xhigh, one point above Carbon. The same evaluation measured 21.5 minutes per task and 283 million total output tokens, so the strongest setting does not automatically provide the best operating point.[5] Teams should compare Neon at high and xhigh against their current model on completed-task quality, total billed tokens, elapsed time, intervention rate and unsupported claims before moving production traffic.

Released
November 18, 2026
Model ID
gemini-4-neon
Unified context
1,048,576 tokens
Standard API
$3 input · $15 output / 1M

What Google released

Google introduced Gemini 4 Neon on November 18 [1] across the paid Gemini API, AI Studio, Gemini for business, Antigravity and paid Gemini app plans. GitHub also made Neon generally available in Copilot for Pro+, Max, Business and Enterprise users.[6]

The model accepts text, images, audio and video and returns text. Google gives it a unified 1,048,576-token allowance that includes input, thinking, function calls and the final answer.[3]

A foreground response can use up to 65,536 tokens. Supported background or continuation flows can extend output within the same one-million-token allowance.[3]

Neon fills the gap between Hydrogen and Carbon

Hydrogen remains the lower-cost execution model. Carbon remains Google's choice for the hardest long-horizon work. Neon targets complex coding and analysis that teams run frequently, where a smaller bill or shorter run than Carbon may matter across many tasks.[1]

Google adapted Neon from Carbon and gave it low, medium, high and xhigh reasoning levels.[2][3] High is the default. That makes effort selection part of the deployment decision rather than an optional tuning detail.

Neon is proprietary, and Google has not disclosed its parameter count.[2][5] Google distributes Neon through hosted products and APIs, with no public weights offered for inspection or self-hosting.

Independent measurements complicate the speed claim

Artificial Analysis measured Neon independently [5] and scored xhigh at 61 on its Intelligence Index. That was one point above Carbon's listed score and seven points above Hydrogen's highest listed score on the evaluator's model pages.

The xhigh run took 21.5 minutes per task, compared with a 12.1-minute median among the models Artificial Analysis grouped with it. Time to the first answer token was 56.51 seconds.[5] Google's description of fast end-to-end completion therefore needs to be tested on each workload and reasoning level.

Artificial Analysis recorded 283 million output tokens across its xhigh Intelligence Index run, compared with a 65.1-million median for comparable models.[5] The model can be fast at generating tokens while still taking longer and producing more of them.

The unified context window changes cost planning

Google's API pricing page [4] lists $3 per million input tokens, $15 per million output tokens and $0.15 per million cached input tokens. Neon has no free API tier, and output pricing includes thinking tokens.

Input, hidden reasoning, function calls and the final response share one limit.[3] A long prompt leaves less room for thinking and output, while a high reasoning setting can raise both cost and latency without changing the visible prompt.

Implicit caching applies to eligible shared prefixes, but explicit context caching is unsupported.[3] Stable instructions should stay early and identical across requests if a team wants to test whether implicit cache hits lower real cost.

Safety evidence is useful but incomplete at launch

Google reports a 0.09% misaligned-action rate, 99.2% on its agent-honesty evaluation and no successful attacks in the evaluated Gray Swan indirect prompt-injection set when real-time safeguards and the highest reasoning level were active.[1] These results describe named tests under Google's setup, not every tool, repository or document an agent may encounter.

The launch model card [2] says the full Safety and Security Report was not available at launch and would follow in the coming days. It points readers to Carbon's report for evaluation methods and policies.

GitHub says administrators can control access through Copilot model policy, and rollout is gradual.[6] Product controls differ by surface, so teams should verify the actual permissions, retention and monitoring in the product they plan to use.

A practical Gemini 4 Neon evaluation

  • Select 20 to 50 representative tasks with known acceptance tests, including multi-file changes, tool failures, ambiguous requirements and long-document analysis.
  • Run the same tasks on the current production model and on Neon at high. Add xhigh only for tasks where high misses a stated acceptance criterion.
  • Record task success, unsupported claims, unnecessary file changes, reviewer interventions, elapsed time, input tokens, output tokens and total billed cost.
  • Keep the harness, tools, repository snapshot, prompts and retry policy fixed. A model comparison is not useful when the surrounding system changes at the same time.
  • Seed documents and tool output with benign conflicting instructions. Confirm that Neon preserves the task boundary, reports the conflict and does not reveal test secrets.
  • Test context limits with realistic files. Include reasoning, tool schemas and expected output in the token budget instead of treating one million tokens as prompt capacity.
  • Adopt Neon only for the task classes where it improves the chosen service objective. Route simple work to a cheaper model and retain Carbon where its result quality justifies the difference.

What remains uncertain

  • Google had not published Neon's full Safety and Security Report when the model card appeared.[2]
  • Most launch benchmark results are cited by Google. Artificial Analysis provides independent general capability, speed, token-use and price measurements, but no evaluation predicts performance on every private workload.[1][5]
  • Google reports lower cost per task than Carbon at high, while Artificial Analysis calls xhigh somewhat expensive among similarly priced models.[1][5] Different settings and task mixes can produce both findings.
  • The cyber variant is limited to trusted defenders through Fairwind and is separate from ordinary API access.[1]
  • Availability, limits and controls vary across the Gemini API, enterprise platform, Gemini app, Antigravity and GitHub Copilot.[1][6]

Frequently asked questions

  • What does Gemini 4 Neon cost? Standard Gemini API pricing is $3 per million input tokens, $15 per million output tokens and $0.15 per million cached input tokens.[4]
  • Is Neon faster than Carbon? Google reports faster end-to-end work in its comparisons. Artificial Analysis found xhigh slow per task, so teams should compare the reasoning levels on their own workloads.[1][5]
  • Should Hydrogen users migrate? Test complex tasks first. Hydrogen remains cheaper and may still fit routine, latency-sensitive execution.[1][4]
  • Should Carbon users migrate? Neon is a candidate for repeated complex work. Keep Carbon where a matched evaluation shows that its deeper planning improves accepted results.[1]
  • Can developers self-host Neon? No public weights are available. Neon is distributed through Google's products and APIs and selected downstream providers.[2][6]
  • Does one million tokens mean one million input tokens? No. Input, thinking, function calls and output share the unified allowance.[3]

Keep learning on Learnetto

Sources

[1] Google: Gemini 4 Neon announcement

[2] Google DeepMind: Gemini 4 Neon model card

[3] Google Cloud: Gemini 4 Neon developer guide

[4] Google AI Studio: Gemini API pricing

[5] Artificial Analysis: Gemini 4 Neon independent measurements

[6] GitHub: Gemini 4 Neon in Copilot