
Quick Answer
Google’s Gemini 4 Neon gives business teams a model designed for long-running coding, research and enterprise workflows, with a one-million-token context window and four reasoning levels. Google says it matched or exceeded the larger Gemini 4 Carbon model in early internal quality tests while completing tasks three times faster, while independent listings provide separate pricing, speed and benchmark observations.
In This Article
Launch Status and Source Map
The available information separates Google’s launch claims from third-party measurements. OpenRouter lists a November 18, 2026 release date and identifies Google AI Studio and Google Vertex as hosting providers. BenchLM lists the same release date, while its model data reflects November 21, 2026.
- Product facts: OpenRouter lists a one-million-token context window, reasoning controls labeled low, medium, high and xhigh, and standard endpoint prices of $3 per million input tokens and $15 per million output tokens.
- Vendor-reported claims: Google reported three-times-faster task completion than Carbon at comparable quality, a 32% lower cost per task at the default high thinking level, and its lowest fabrication rate in internal Google Antigravity testing.
- Independent observations: OpenRouter reports provider-level latency, throughput and uptime for Google AI Studio and Google Vertex. BenchLM records 16 source-linked benchmark rows and 16 of 952 tracked benchmark slots with displayable evidence.
- Secondary source: Detailed claims about multimodal inputs and a one-million-token context window came from Interesting Engineering. Equivalent primary documentation was not available in the supplied research.
Confirmed Launch Facts
Gemini 4 Neon is a fast, mid-priced model in Google’s Gemini 4 family, positioned between Gemini 4 Hydrogen and Gemini 4 Carbon. OpenRouter describes it as a model for repository and terminal coding, web builds, multi-turn tool use and multimodal STEM problems.
Its one-million-token context window allows it to process large volumes of information across spreadsheets, PDFs, charts, audio and video. OpenRouter lists reasoning controls labeled low, medium, high and xhigh, and says Neon supports long decodes of up to one million output tokens through continuation.
OpenRouter also lists Google AI Studio and Google Vertex as available providers, so access depends on the provider and its current routing configuration.
Google-Reported Performance
Google says Neon matched or exceeded Carbon in overall quality during early internal evaluations while running three times faster. These comparisons were prepared by Google and were not independently verified in the supplied material.
Google also reported a 32% lower cost per task than Carbon at the default high reasoning level. Its estimate of roughly 50% to 80% savings against leading competing models applied only to selected evaluations.
For coding, Google reported a first-place score of 58.6 on the Artificial Analysis Coding Agent Index. It reported 65.2% on Terminal-Bench 4.1 and 39% on SWE-sweep when tested in Google Antigravity, compared with 12.8% on the mini-SWE harness. Other reported results included 70.5% on Terminal-Bench Science, 84% on RiemannBench, and reported placements on Legal Research Bench, Tax Agent Bench and Zapier’s AutomationBench.
Google’s reported Intelligence Index score was 61 at the xhigh level, placing it third at the time. The model also ranked second on the Vals Index and the Agent Arena Index for steerability.
Independent Benchmark Context
BenchLM’s November 21, 2026 model profile gave Gemini 4 Neon a score of 91.3 out of 100 and ranked it second among 215 models. Within BenchLM’s tracked catalog, the model ranked first in coding, third in agentic tasks, second in reasoning and math, and fifth in multimodal work. Those positions apply to BenchLM’s category weights and available benchmark rows, not to every business workload.
BenchLM identified coding as its strongest eligible category and multimodal grounded work as its lowest eligible category. The profile said 16 of 952 tracked benchmark slots had displayable evidence, so its overall score sat on a selective evidence base. It also listed 110 tokens per second for speed and 56.51 seconds to the first token.
OpenRouter’s provider data showed materially different operating conditions across routes. Google AI Studio Standard reported 5.10-second latency, 74 tokens per second and 96.92% uptime. Google Vertex Standard reported 3.52-second latency, 78 tokens per second and 98.35% uptime. Priority and Flex routes reported different prices, latency and throughput, which makes the selected route part of any workflow comparison.
Pricing and Availability
The standard prices listed by OpenRouter were $3 per million input tokens and $15 per million output tokens, with cache reads at $0.15 per million. Check current provider listings before changing a production workflow.
OpenRouter also listed lower standard-route prices for Google AI Studio Flex and higher prices for priority routes. Its weighted customer-price section reported an effective input price of $0.4945 per million and an effective output price of $14.26 per million for the listed selection. Provider mix, caching and routing can change the amount actually paid.
Availability was listed through Google AI Studio and Google Vertex.
What Business Teams Should Evaluate
Compare Neon with the model used in the current workflow on one representative task. The comparison framework should cover the task definition, quality criteria, completion latency, total input and output tokens, reasoning setting, human-review time and source-citation accuracy.
- Representative task: Select a workflow that resembles the actual production work.
- Quality criteria: Record the expected output and the checks used to assess it.
- Latency: Measure completion time and time to the first token.
- Token use: Record input, output and cached-token consumption.
- Reasoning: Test the reasoning level required by the workload.
- Review effort: Record the time required for human verification.
- Citations: Check whether responses provide verifiable source references.
Before Changing Production Workflows
- Confirm current model access and availability in the official Google documentation.
- Record a baseline for output quality, latency and review time.
- Use a representative, non-sensitive workflow for the evaluation.
- Check whether outputs include verifiable citations.
- Measure total token consumption and the selected provider route.
- Require human review for consequential outputs.
Claude 3.5 Sonnet Upgrade and Mistral NeMo Release
In a 2024 release, Anthropic reported that the Claude 3.5 Sonnet upgrade improved SWE-bench Verified performance from 33.4% to 49.0%. Anthropic also introduced a computer-use beta through the Anthropic API, Amazon Bedrock and Google Cloud’s Vertex AI.
Mistral’s 2024 NeMo release included pre-trained base and instruction-tuned checkpoints under the Apache 2.0 license. Mistral said the model offered state-of-the-art reasoning, world knowledge and coding accuracy within its size category.
Key Considerations
- Google’s quality, speed and cost comparisons with Carbon came from early internal evaluations.
- The reported 50% to 80% savings against competing models applied to selected evaluations.
- Compare reasoning settings on the workload before choosing one.
- Google’s fabrication-rate result came from internal Google Antigravity testing.
- Third-party benchmark coverage is selective, and provider routing can change latency, throughput and price.
FAQ
What Is Gemini 4 Neon?
Gemini 4 Neon is a Google model for complex coding, scientific research and enterprise workflows. Its one-million-token context window allows it to process large volumes of information across spreadsheets, PDFs, charts, audio and video.
How Fast Is Gemini 4 Neon?
According to Google, Neon matched or exceeded Carbon in overall quality during early internal evaluations while completing tasks three times faster.
How Much Does Gemini 4 Neon Cost?
The standard prices listed by OpenRouter were $3 per million input tokens and $15 per million output tokens. Cached input tokens were listed at $0.15 per million. Provider and routing prices differed.
What Is Gemini 4 Neon’s Context Window?
OpenRouter lists a one-million-token context window and says Neon supports long decodes of up to one million output tokens through continuation.
What Can Gemini 4 Neon Do?
Google positions Neon for repository-level debugging, multi-file refactoring, scientific research, legal and tax analysis, enterprise data analysis and cybersecurity work. It can also direct parallel Hydrogen subagents within Google Antigravity.
Related Coverage
- Google I/O 2026: Vibe Coding Comes to Your Phone – BizScoreAI
- Google’s Frozen v2 Chip Aims to Make Gemini More Efficient
- Google Research on AI Spam Detection Explained
Sources
See our editorial disclaimer for how our articles are produced.
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.