AI Models & Releases
Gemini 4 Neon Explained: What It Is and Who It Is For
Google's Gemini 4 Neon is a new mid-tier AI model that aims to deliver near-flagship reasoning quality at faster speeds and lower cost than the top-of-the-range Gemini 4 Carbon. It quietly completes the Gemini 4 family and reshapes who can access frontier-level AI.
Key takeaways
- Gemini 4 Neon is Google's mid-tier Gemini 4 model, sitting between Hydrogen (fast, everyday) and Carbon (the most powerful), targeting complex coding, enterprise analysis, and science workflows.
- It is the first Gemini 4 frontier model available to Google AI Pro subscribers, not just Ultra, significantly broadening access to near-flagship reasoning.
- Four adjustable thinking levels (Low, Medium, High, Extra High) let users and developers trade reasoning depth against latency and token cost on every request.
- In Google Antigravity, Neon can orchestrate parallel Gemini 4 Hydrogen subagents, acting as a planner while Hydrogen handles routine execution steps.
- Neon is priced at $3 per million input tokens and $15 per million output tokens, with Google claiming a 32% lower cost per task at the default High setting compared with Carbon.
Google has shipped four Gemini 4 models in roughly five weeks, and the latest one is not just a gap-filler. Gemini 4 Neon lands as the model that most AI Pro subscribers will actually reach for day to day: powerful enough for serious coding and research, fast enough to stay out of your way, and priced at a tier that does not require a premium Ultra subscription. For anyone watching the AI model wars, it is also a signal that Google is now playing the full lineup game, with a distinct model for every rung of the performance-cost ladder.
What Is Gemini 4 Neon?
Gemini 4 Neon is a frontier-class large language model from Google, announced and released on November 18, 2026. Google describes it as built for complex software engineering, knowledge work, and science workflows. The model is the fourth member of the Gemini 4 family, joining Gemini 4 Argon, Gemini 4 Carbon, and Gemini 4 Hydrogen. According to Google, it brings quality close to Gemini 4 Carbon on coding, reasoning, and agentic tasks, at what Google calls Flash-tier speed and cost.
In the Gemini app's model picker, Neon appears under the label "Deeper analysis," sitting between "Everyday help" (Hydrogen) and "Code and complex work" (Carbon). It also replaces Gemini 4 Argon and retires Gemini 3.1 Pro from the app entirely, tidying a lineup that had grown cluttered.
Where Neon Fits in the Gemini 4 Family
The Gemini 4 naming scheme follows the periodic table, and understanding the family tree is essential for making sense of Neon.
Google frames the three main tiers this way: Carbon handles the hardest, longest-running projects; Hydrogen covers fast everyday execution; and Neon balances depth and efficiency between them. In practical terms, Neon is for the work that is too demanding for Hydrogen but does not need the full weight of Carbon.
One immediately useful consequence of that positioning is access. According to some sources, Gemini 4 Carbon requires a higher-tier subscription in the Gemini app, while Neon is available to both AI Pro and AI Ultra members. That access expansion means a much larger share of paying users can get a frontier-calibre model in the Gemini app without upgrading.
Technical Specs: Context, Modalities, and Thinking Levels
Neon's headline specification is a context window of approximately one million tokens, shared across input, reasoning, and output. That window is large enough to ingest sprawling codebases, long documents, charts, audio, and video in a single session.
On the input side, the model accepts text, images, video, audio, and PDFs, and returns text output. Image generation, audio generation, and the Gemini Live API are not supported on Neon; those features rely on separate models.
One of Neon's most distinctive engineering choices is its four-level thinking system: Low, Medium, High (the default), and Extra High (xhigh). The setting lets users and developers trade reasoning depth against latency and token consumption on a per-request basis. Google recommends evaluating all four levels and choosing the lowest one that meets a workload's quality requirement, reserving xhigh for difficult multi-step problems where the high setting falls short. Notably, while Gemini 4 Carbon and Argon default to medium and top out at high, Neon starts at high and extends all the way to xhigh, giving it more head room for complex reasoning tasks.
The model also supports computer use, code execution, file search, function calling, structured outputs, and grounding through Google Search and Google Maps.
What Neon Is Built to Do Well
Google's positioning centres on three major use-case clusters.
The first is software engineering. Neon is designed for multi-file debugging, refactoring, and what Google calls surgical, targeted code edits. In coding-specific benchmarks from Google's launch table, Neon scores 65.4% on CursorBench 4.1 and 65.2% on Terminal-Bench 4.1. Google also reports that Neon ranks first on the Artificial Analysis Coding Agent Index with a score of 58.6. On the safety side, it matched Gemini 4 Carbon with a 0.0% cheating rate on Terminal-Bench 4.0, and recorded the lowest cheating rate of any Gemini 4 model on a separate coding evaluation.
The second cluster is knowledge work and enterprise analysis. Neon can synthesize spreadsheets, PDFs, charts, audio, and video within that one-million-token context window and produce structured deliverables. On the Vals Index, it ranks second with a score of 70.9%, and it reaches first place on both the Legal Research Bench and the Tax Agent Bench, suggesting particular strength in document-heavy professional workflows.
The third cluster is science. Neon scores 70.5% on Terminal-Bench Science and 84.0% on RiemannBench, and it places first on HLE-Diamond at 71.3% according to Google's launch benchmarks. Google's own demo showed the model using its multimodal and spatial-reasoning capabilities to process a 3D scan and reconstruct a full anatomical skeleton, illustrating how those capabilities can combine in a single long-running task.
Across all three areas, Google also highlights two reliability improvements: strong multi-step instruction following, and what it describes as the lowest fabrication rate of any Gemini model in internal Google Antigravity testing to date.
The Steerability Angle
One of the less obvious but practically important claims Google makes about Neon is its steerability score. On the Agent Arena Index, Neon ranks second for steerability, which Google defines as a model's ability to adapt cleanly to user corrections mid-task. For long-running agentic workflows, that matters a great deal: a model that can absorb a course correction without derailing the entire task is far more useful than one that requires users to restart from scratch.
Neon is also described as trained to complete multi-step work without cutting corners, fabricating results, or overstepping user-defined boundaries. Google's internal monitoring for chain-of-thought and action misalignment showed alert rates below the already-low baselines set by Carbon, suggesting improved behavioural alignment.
Neon as an Agent Orchestrator
Beyond single-session use, Google is positioning Neon as the planning brain in multi-agent pipelines. Inside Google Antigravity, a new Orchestrator feature lets Neon act as the lead model while delegating routine steps to parallel Gemini 4 Hydrogen subagents. The workflow works like this: Neon handles upfront planning, architecture decisions, and complex reasoning, then hands off routine code edits, package updates, and test runs to Hydrogen. If Hydrogen encounters a problem it cannot resolve, control passes back to Neon to inspect the state, repair the plan, and resume execution.
This Neon-plus-Hydrogen pairing is designed to keep agent loops cost-efficient. Because Hydrogen is a lighter, faster, cheaper model, offloading predictable steps to it means the expensive Neon thinking budget is reserved for the parts of a task that genuinely need it.
Pricing and Access
Neon is priced at $3 per million input tokens and $15 per million output tokens at launch, with a steep discount on cached input tokens. That makes it twice the cost of Gemini 4 Hydrogen (priced at $1.50 input and $7.50 output) and, at launch, slightly more expensive than Gemini 4 Carbon and Argon, which are on an introductory rate that is set to rise.
Google says that at Neon's default High thinking level, the cost per task is 32% lower than Carbon's on its benchmark comparison. The qualifier matters: that figure is a task-level comparison on specific evaluations, not a simple per-token calculation. Actual costs in production will depend on how much reasoning each job triggers.
For consumer access, Neon is available in the Gemini app for Google AI Pro and Ultra subscribers. For developers, it is available via the paid Gemini API in Google AI Studio and Android Studio, and via the Google Cloud Vertex AI platform. Enterprise users can reach it through Gemini for business via the API or the model dropdown. There is no free API tier for Neon.
Benchmarks in Context
At its xhigh thinking level, Neon scores 61 on the Artificial Analysis Intelligence Index, placing it third overall at launch. Google's benchmark table also shows Neon scoring 56.3% on AutomationBench and 75.8% on Agents' Last Exam, results that Google reports as competitive with other frontier models it names in the comparison. On the Agent Arena Index overall, Neon ranks second.
It is worth noting how to read these numbers honestly. Benchmark results are always measured at a specific thinking level and under specific testing conditions. The xhigh scores that headline Google's intelligence rankings use the most expensive reasoning mode; the same model at Low or Medium thinking will produce different results and consume far fewer tokens. Google itself recommends using the lowest setting that meets a workload's needs, which means real-world performance will vary considerably depending on how the model is configured.
What to Watch Next
Several near-term developments are worth tracking as Neon settles into the market.
First, Gemini 4 Carbon and Argon are currently on an introductory API price that is scheduled to rise on December 1, 2026. Once that happens, the relative cost positioning of Neon versus Carbon may shift, and the "32% cheaper per task" comparisons Google uses will need to be re-evaluated against new Carbon rates.
Second, Neon's rollout to Gemini app users is staged, so not every AI Pro subscriber will see it in their model selector on day one. Watch for confirmation that the rollout is complete.
Third, Google has described a Gemini 4 Krypton variant that some testers reportedly spotted before Neon launched. Whether that becomes the next named model in the elemental series, or represents something else in the pipeline, is an open question.
Finally, the Orchestrator pairing of Neon and Hydrogen inside Google Antigravity is a relatively new agent-workflow pattern. How that plays out in real enterprise deployments, and whether it becomes the standard way professional teams structure long-running tasks, will tell us a lot about whether the multi-model strategy Google is pursuing actually delivers on its efficiency promises.
Sources
- Introducing Gemini 4 Neon – Google Blog
- Google launches Gemini 4 Neon with Carbon quality at Flash speeds, live for AI Pro – 9to5Google
- Gemini 4 Neon launches for Google AI Ultra and AI Pro plans – Android Authority
- Google launches Gemini 4 Neon with Opus 5.5-level coding performance – VentureBeat
- Google ships Gemini 4 Neon, its fourth new Gemini model in a month – The Decoder
- Google Launches Gemini 4 Neon for Faster Coding and Analysis – iClarified
- Google Launches Gemini 4 Neon for Coding, Knowledge Work and Science – Unite.AI
- Gemini 4 Neon in Google Antigravity – Antigravity Blog
- Developer's guide to Gemini 4 Neon – Google Cloud Documentation
- Gemini 4 Neon – OpenRouter API Pricing and Providers
- Gemini 4 Neon: Intelligence, Performance and Price – Artificial Analysis
- Google launches Gemini 4 Neon on Gemini, Antigravity, and APIs – Testing Catalog
- Gemini 4 Neon Beats Claude Opus 5.5, GPT 6.2 Sol On Most Benchmarks – OfficeChai
- Google AI on X – Gemini 4 Neon announcement thread
- Gemini 4 Neon benchmarks, specs, pricing and comparisons – Kingy.ai
Frequently asked questions
