Google adds a fourth model to the Gemini 4 family

Google released Gemini 4 Neon on November 18, 2026, as the fourth model in its Gemini 4 lineup, following Argon, Carbon, and Hydrogen.

The company describes it as a model that brings "frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows."

The Verge reported that the new model sits between Carbon and Hydrogen in the lineup. While Carbon handles the hardest, longest-running projects and Hydrogen manages fast everyday tasks, Neon balances depth and efficiency for work that needs both.

Gemini 4 Neon is priced at $3 per million input tokens and $15 per million output tokens, with cached input tokens available at $0.15 per million, a 95 percent discount.

The model supports a 1 million-token context window and began rolling out on November 18 to developers through the Gemini API in Google AI Studio, to enterprise customers through Gemini for business, and to Pro and Ultra subscribers in the Gemini app and Google Antigravity.

Benchmark performance across coding, science, and analysis

Neon achieved strong results across independent evaluations.

It ranks second globally with a score of 70.9 percent on the Vals Index and third overall with a score of 61 on the Artificial Analysis Intelligence Index at its highest thinking level, according to Google's announcement.

On the Agent Arena Index, a new benchmark measuring real-world agent sessions, Neon ranked second on steerability, adapting cleanly to mid-task user corrections.

On the coding front, the company said Neon achieved the top score on the Artificial Analysis Coding Agent Index at 58.6, with a score of 65.2 percent on Terminal-Bench 4.1 and 46.2 percent on FrontierCode 2.0.

For production-ready code, Neon handles multi-file debugging and refactoring across repositories with targeted edits. On the SWE-sweep benchmark, it scored 12.8 percent on the mini-SWE harness and reached 39.0 percent in Google Antigravity, showing what the company's agentic runtime unlocks.

In scientific and mathematical reasoning, the model was best in class on Terminal-Bench Science with 70.5 percent, RiemannBench with 84.0 percent, and HLE-Diamond with 71.3 percent. It also ranked first on the Legal Research Bench and Tax Agent Bench.

Google said the model debuts second globally with a score of 1,336 on the Agent Arena Index across more than 4,200 real-world agent sessions.

Enterprise and analytical capabilities

Google positioned Neon for enterprise workloads that involve spreadsheets, PDFs, charts, audio, and video within a single 1 million-token context window. The company said the model can synthesize across all these inputs and turn them into structured deliverables.

Independent evaluations showed Neon ranking first on most benchmarks across finance, accounting, spreadsheet modeling, and complex constraint following on Surge AI.

It scored 73.5 percent on Chartography for reading dense professional charts and performed on par with leading models on Agents' Last Exam v1.1 with 75.8 percent.

The model achieved the top score on Zapier's AutomationBench at 56.3 percent and on Andon Labs' Vending-Bench 2, a long-horizon business simulation, where it earned $16,843.

The company also highlighted Neon's strength in multimodal understanding. Because it can process spreadsheets, PDFs, charts, audio, and video in a single context, Google said it is suited for knowledge workers who need to cross-reference information across formats.

The model ranks second on the Vals Index across multiple categories and sits on the Pareto frontier for both quality and cost.

Agentic workflow and cybersecurity features

Google designed Neon for long-running agent workflows, sustaining performance through complex, tool-heavy tasks. The company said the model can run standalone or act as an orchestrator directing parallel Hydrogen subagents in Google Antigravity.

It cuts down on redundant file reads, shares progress as it works, and handles context compaction reliably in long sessions. Neon uses four thinking levels that scale steadily, letting teams dial intelligence and token spend to what each workload needs.

The company said the model's architecture is built for production-grade enterprise deployment with reliable context compaction.

A dedicated cybersecurity variant, Gemini 4 Neon for cyber, is rolling out to trusted defenders through the Fairwind Program.

Google said the cyber version produces more secure code than Carbon across all difficulty levels on AutoBaxBench, achieving a state-of-the-art score of 76 percent on the hard tier.

On WebExploitBench v1.2, it reached 51.26 percent at half the cost of competing large models.

The company also highlighted Neon's vulnerability-finding performance on SEC-bench Pro at 90.5 percent, its exploitation validation on ExploitGym, and its patching capability on Collinear's CWE-bench v2 at 70.7 percent, a 3.4 percent improvement over Carbon.

Google said the cyber version is designed for the full code hardening lifecycle: finding vulnerabilities, validating risk through exploitation, and patching the root cause.

Safety, alignment, and pricing position

Google said Neon maintains Carbon-level protections against severe misuse and prompt injection while declining far less legitimate work in cybersecurity and biosecurity areas. The model refused just 1.98 percent of tasks on Terminal-Bench 4.0 and zero percent on Terminal-Bench-Science.

On the company's agentic safety suite, it achieved a 0.09 percent misaligned action rate and a 99.2 percent pass rate on agent honesty evaluations.

On Gray Swan's indirect prompt injection benchmark, Neon resisted every attack at its highest thinking level, matching the 0.0 percent attack success rate of other Gemini models.

The company also reported that internal monitoring showed alert rates from chain-of-thought and action misalignment monitors remained below Carbon's already low baselines, suggesting improved alignment propensity.

At its default high thinking level, Google said Neon pairs a high Artificial Analysis Intelligence Index score with a 32 percent lower cost per task than Carbon.

At the highest thinking level, the company said the model costs 36 percent or more less per task than competing frontier models at their highest setting.

On both the Vals Index and Agent Arena at the high setting, Neon sits on the Pareto frontier with roughly 50 to 80 percent lower cost per task than top competitors.

Common questions

What is Gemini 4 Neon?

Gemini 4 Neon is the fourth model in Google's Gemini 4 family, announced on November 18, 2026. It is designed for complex coding, analytical reasoning, and production-grade enterprise workflows, sitting between Gemini 4 Carbon and Gemini 4 Hydrogen in capability.

How much does Gemini 4 Neon cost?

Gemini 4 Neon costs $3 per million input tokens and $15 per million output tokens, with a 95 percent discount to $0.15 per million cached input tokens. It has a 1 million-token context window.

How does Gemini 4 Neon compare to other models?

Neon ranks second on the Vals Index (70.9 percent) and third on the Artificial Analysis Intelligence Index at its highest thinking level. It also leads on several benchmarks including the Legal Research Bench, Tax Agent Bench, and Terminal-Bench Science.

Sources & further reading

  1. Introducing Gemini 4 Neon - Google Blog ↗
  2. Google released another Gemini 4 model - The Verge ↗
How this story was made

Written by AI Lookout with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.

Our editorial standards
Back to all stories