Home / Blog / Google introduces Gemini 4 Neon: hero image for Gemini 4…
Tech News

Google introduces Gemini 4 Neon: hero image for Gemini 4 Neon

Google launched Gemini 4 Neon with a 1 million-token context window, ranking third on the Artificial Analysis index while cutting task costs by 32 percent.

By Dillip Chowdary • Nov 19, 2026 • Source: Google Keyword Blog

Google introduces Gemini 4 Neon: hero image for Gemini 4 Neon
Image: Google Keyword Blog

At a glance

Product name
Gemini 4 Neon
Release date
November 18, 2026
Base pricing
$3 input and $15 output per million tokens
Cached pricing
$0.15 per million tokens (95% discount)
Context window
1 million tokens
Platform availability
Google AI Studio, Gemini app, Google Antigravity

Google launched Gemini 4 Neon on November 18, 2026, introducing a frontier-class artificial intelligence model designed for fast end-to-end task completion across software engineering, enterprise analysis, and scientific research. According to Google Keyword Blog's report, the system ranks third overall on the Artificial Analysis Intelligence Index with an intelligence score of 61 at its xhigh thinking tier. Google priced the release at $3 per million input tokens, $0.15 per million cached input tokens, and $15 per million output tokens within a 1 million-token context window.

This analysis breaks down the benchmark results, pricing structure, autonomous agent capabilities, cyber defense features, and safety guardrails across production environments. Enterprise software teams, data scientists, and systems administrators can evaluate how the model handles long-horizon coding tasks, multi-step instruction compliance, and large document synthesis. The coverage examines verified performance across independent harnesses to assist technical leaders assessing migration from earlier releases.

01Enterprise spreadsheets and scientific benchmark evaluations

Across corporate analysis, the model processes text, spreadsheets, dense PDFs, charts, audio, and video inside its 1 million-token window to deliver structured outputs. On the Vals Index, it captured the second spot with a 70.9 percent score while clinching first place on both the Legal Research Bench and Tax Agent Bench. Independent testing by Surge AI positioned the model in first place across finance, accounting, spreadsheet modeling, and complex constraint following.

In visual analysis, the model achieved 73.5 percent on Chartography for parsing professional charts and scored 75.8 percent on Agents' Last Exam v1.1. In simulated operations, it secured first overall on Zapier's AutomationBench at 56.3 percent and generated $16,843 on Andon Labs' Vending-Bench 2 business simulation. On the Agent Arena Index, the architecture debuted in second place globally with a score of 1336 across more than 4,200 real-world sessions.

For advanced quantitative research, the model recorded top-tier results across complex scientific benchmarks. It led the field on Terminal-Bench Science with a 70.5 percent accuracy mark while recording 84.0 percent on RiemannBench mathematics problems. On the HLE-Diamond evaluation suite, the model registered 71.3 percent, demonstrating consistent analytical reasoning across rigorous technical disciplines.

02Multi-file repository debugging and agentic coding

Engineered for autonomous execution, the model directs multi-file debugging and surgical repository refactoring without unnecessary file reads. Operating as an orchestrator in Google Antigravity, it coordinates parallel Hydrogen subagents, shares live progress telemetry, and compacts session contexts reliably during long runs. The system earned first place on the Artificial Analysis Coding Agent Index with a 58.6 score, alongside 65.2 percent on Terminal-Bench 4.1 and 46.2 percent on FrontierCode 2.0 at default settings.

Benchmark testing on the mini-SWE harness showed a 12.8 percent baseline mark that expanded to 39.0 percent when deployed inside the Google Antigravity environment. In practical demonstrations using the /swarm command, the system built a WebGL2 ski simulator spanning 10,200 custom shader lines on Mont Blanc elevation maps with zero pre-made assets. Another swarm implementation transformed a single photograph into a 100,000-cell Google Sheet powered by 500,000 active formulas calculating depth maps and scene relighting.

Additional builds featured a 3D historical reconstruction of Alan Turing's Cambridge study with an operational Universal Turing Machine alongside a Blender MCP laboratory visualization. The model also ranked second for steerability on the Agent Arena Index by absorbing mid-task user corrections cleanly. Internal Antigravity testing revealed the model achieved the lowest fabrication rate recorded among Google's production agent offerings.

03Four thinking tiers and token pricing

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

Google established four graduated thinking tiers to let engineering organizations balance operational latency and computational expense against workload complexity. At the default high thinking level, the model operates on the Pareto frontier across the Vals Index and Agent Arena. It provides enterprise teams with a 32 percent cost savings per completed task compared to Carbon, while running 50 to 80 percent below competing frontier models.

For demanding scientific and mathematical challenges, the xhigh thinking configuration unlocks deeper chain-of-thought calculation to achieve peak reasoning scores. Operating at this top level, the model maintains a 36 percent cost advantage over competing foundation models evaluated at their maximum configurations. Input caching delivers a 95 percent discount, dropping the base $3 per million input rate down to $0.15 per million cached tokens.

| Configuration tier | Relative cost efficiency | Target application | | High thinking | 32% below Carbon; 50% to 80% below frontier models | Standard enterprise tools and routine coding | | xHigh thinking | 36%+ below frontier models at peak settings | Deep mathematical research and complex engineering |

Note: Rollout began immediately across the Gemini API in Google AI Studio and Gemini for business model menus. Paid Google AI Pro and Ultra subscribers can access the model directly inside the Gemini mobile application and the Google Antigravity workspace platform for developer operations.

04Automated penetration testing and vulnerability patching

The security-focused edition was designed to generate hardened source code while proactively detecting system vulnerabilities before malicious actors exploit them. On AutoBaxBench, it surpassed Carbon across easy, medium, and hard benchmarks, achieving a top score of 76 percent on the hard tier. Web penetration evaluations on WebExploitBench v1.2 placed the model at a state-of-the-art 51.26 percent score at half the cost of rival systems.

  1. Finding code vulnerabilities SEC-bench Pro reaches 90.5 percent accuracy while cutting cost per challenge by approximately half.
  2. Validating risk through exploitation ExploitGym matches frontier performance using half the tokens consumed by Carbon.
  3. Patching vulnerabilities Collinear CWE-bench v2 delivers a 70.7 percent resolution score, marking a 3.4 percent gain over Carbon.

Defenders utilize these coordinated lifecycle steps to inspect source trees, confirm exploitability paths, and commit automated code patches before deployment. By halving token consumption during threat validation, operations teams can evaluate larger vulnerability queues without expanding compute budgets. Early access for the specialized cybersecurity variant will open through the Fairwind Program in the coming days.

05Misuse refusal rates and execution alignment

Safety systems integrate defenses against prompt injection and severe misuse derived from the Gemini 4 infrastructure stack, matching protections established in Carbon. The model enforces real-time safeguards against biosecurity hazards and cyber exploitation while lowering refusal rates on legitimate enterprise operations. System evaluations showed a task refusal rate of 1.98 percent on Terminal-Bench 4.0 and zero percent refusals on Terminal-Bench-Science tasks.

Google deployed dedicated chain-of-thought and action misalignment monitors to track execution across internal developer environments and production API surfaces. Alert frequencies remained below the low baseline thresholds set by Carbon, reflecting improved alignment propensity during complex tool manipulation. The architecture is trained to avoid fabricating data, cutting corners on multi-step instructions, or exceeding authorized agent permissions.

Frequently asked questions

What is the token pricing for Gemini 4 Neon?

Base pricing is set at $3 per million input tokens and $15 per million output tokens. Cached input tokens cost $0.15 per million, reflecting a 95 percent discount.

Where is Gemini 4 Neon available?

The model is available in the Gemini API via Google AI Studio, Gemini for business, and for Google AI Pro and Ultra subscribers in Google Antigravity and the Gemini app. The cyber defense variant is rolling out to Fairwind Program participants.

How does Gemini 4 Neon score on key coding and agent benchmarks?

It scored 58.6 on the Artificial Analysis Coding Agent Index, 65.2 percent on Terminal-Bench 4.1, and 46.2 percent on FrontierCode 2.0. In Google Antigravity, its SWE-sweep score reached 39.0 percent.

What context window and thinking levels does the model offer?

The model features a 1 million-token context window and provides four adjustable thinking levels. The xhigh thinking level delivers peak reasoning scores while costing 36 percent less per task than competing models at top settings.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →