Gemini 4 Neon Released: Google's Strongest Coding Model Tops the Rankings

Gemini 4 Neon tops the coding agent index on release and becomes cheaper than Carbon in December, making it Google's best-value frontier coding model.
Google's Gemini 4 family adds Neon, positioned between the speed-focused Hydrogen and the heavyweight Carbon, targeting complex software engineering, knowledge work, and scientific research. Within a day of release, Neon topped the Artificial Analysis Coding Agent Index with 58.6, ranked third on the comprehensive intelligence index with 61, and scored 65.2% on Terminal Bench 4.1. It features a 1M token context and four thinking levels (default high), has no free tier, and is available via the paid Gemini API, Gemini for Business, Anti-Gravity, and Pro/Ultra plans. Priced at $3 input / $15 output per million tokens with a 95% cached-input discount, Neon becomes the cheaper frontier coding model after Carbon's price increase on December 1. The article notes launch-week benchmarks are only early signals and recommends validating with your own workflows before migrating.
Google's Gemini 4 family has a new member. According to analysis from a YouTube channel, Gemini 4 Neon was released on Wednesday, November 18, and within less than a day it climbed to the top of the Artificial Analysis Coding Agent Index. Not second place — first. The channel had previously covered two other models, Hydrogen and Carbon, and Neon's arrival completes the family puzzle. Its pricing arrangement may make it Google's most cost-effective model to date.

Four Models, Each With a Clear Role
Google currently has four Gemini 4 models, each with a well-defined role. Hydrogen focuses on speed, handling daily Q&A and routine coding; Carbon is the heavyweight, taking on the hardest problems; Neon sits in the middle, targeting complex software engineering, knowledge work, and scientific research with frontier-level intelligence.
To use the analogy from the channel's video: Hydrogen is the fast commuter car for daily driving, Carbon is the truck hauling the heaviest loads, and Neon is the sports car — fast enough for everyday driving, strong enough when you get serious. Google DeepMind built Neon to target the middle ground where most developers actually live: code too complex for fast models to handle, but not requiring the most expensive brain in the entire product line.
Benchmarks: Topping the Rankings on Real Coding Tasks
Neon's results aren't just another marketing launch. On the Artificial Analysis Coding Agent Index, it ranks first with a score of 58.6. This index doesn't measure the ability to write a single function — it measures how well a model performs real coding agent work: planning, calling tools, and completing tasks autonomously. On the comprehensive Artificial Analysis Intelligence Index, Neon ranks third with a score of 61 at the highest thinking level. On Terminal Bench 4.1, a benchmark requiring models to run commands in a real terminal and self-correct errors, Neon scores 65.2%.

What these numbers mean: a model can excel on quiz-style tests yet be completely incapable of building software. The benchmarks above test exactly the messy real-world parts — long tasks, real tools, real codebases. Those are the areas where Neon wins.
Artificial Analysis Coding Agent Index is a benchmark maintained by the independent evaluation firm Artificial Analysis, specifically measuring a model's end-to-end capability as a "coding agent": given a real repository and vague requirements, the model must plan steps on its own, read and write files, invoke terminal and testing tools, iterate and fix based on error messages, and ultimately submit runnable code changes. Unlike function-level benchmarks such as HumanEval, it doesn't provide complete function signatures or limit responses to a single turn — it examines tool use, error recovery, and context management over long-horizon tasks. The index typically combines pass rates and efficiency scores across multiple agent scenarios (such as bug fixing, feature addition, and refactoring), so the score better reflects a model's practical usability in IDE/CI environments — which is why Neon's 58.6 is seen as a meaningful signal.
Inside the Model: Three Key Configurations
Neon has three notable features. First is context: it has a 1 million token context window, capable of ingesting an entire codebase, a complete research paper, or hundreds of pages of documents at once — and still remembering the beginning when it reaches the end. Second is thinking levels: Neon offers four thinking levels, defaulting to high. More thinking yields better answers on hard problems but is slower and more expensive; dial it down for simple tasks and up for difficult ones. Third — and this is important — Neon has no free tier. It's a paid model, available through the paid Gemini API in AI Studio, Gemini for Business, Anti-Gravity within the Google AI Pro and Ultra plans, and the Gemini app in the Pro and Ultra plans.
Thinking levels are a reasoning budget control mechanism widely introduced in reasoning models in recent years: before generating a final answer, the model can allocate different amounts of internal reasoning tokens for self-deliberation, verification, and backtracking. Higher levels mean the model tends to engage in multi-step reasoning, try multiple approaches, and self-check, improving accuracy on hard problems but increasing latency and output costs accordingly; lower levels respond faster and cheaper, suitable for simple queries or high-throughput scenarios. Neon defaults to high, meaning it's quality-first out of the box. Developers can switch dynamically based on task difficulty — dialing simple subtasks down to low levels in agent pipelines to save on bills, while reserving high levels for critical decisions.
Pricing Subtlety: Cheaper Than Carbon Starting in December
Neon's pricing is $3 per million input tokens and $15 per million output tokens. Cached input — text that's repeatedly reused — is just 15 cents per million, a 95% discount on repeated context.

The real turning point is the comparison with Carbon. Carbon's introductory pricing is $2 input / $10 output, but only until November 30. Starting December 1, Carbon rises to $4 input / $20 output. From that day on, Neon's per-token cost will be lower than Carbon's. For reference, OpenAI's GPT 6.2 Astra is priced at $8 input / $40 output — more than double Neon's price. If you're building coding agents and watching your API bills, Neon suddenly becomes the cheapest frontier coding model on the market.
Cached input refers to the model caching recurring identical prefixes (such as system prompts, long codebases, or fixed documents), with subsequent requests hitting the cache billed at a steep discount rather than reprocessing all input tokens at full price. This is especially important for coding agents: agents often make multiple calls on the same codebase with highly repetitive context, and cache hit rates can exceed 80%. Neon's cached input price is 15 cents per million tokens, about 5% of its standard $3 input price — a 95% discount. This makes the actual unit cost of long-context, multi-turn agent workflows far lower than the sticker price suggests, and is an important support for its "best value" claim.
How to Choose: Three Scenarios, Three Models
The selection logic is clear. For speed, use Hydrogen — quick answers, routine coding, fast drafts. For the hardest problems in the room, where you need the deepest reasoning Google can offer, use Carbon. And for building real software, use Neon: coding agents, complex refactoring, long multi-step tasks, scientific work — these are Neon's home turf.

It's the balanced choice: frontier intelligence, the top coding agent score, and a price below Carbon starting in December. To be honest, Neon is a brand-new model, and the benchmark data above comes from launch week. Real-world results will take a few weeks to stabilize, but the early signals are already strong. If you're already using Anti-Gravity or the Gemini API, you can try it now.
Editorial View
The four models in the Gemini 4 family form a complete gradient from speed to depth, and Neon fills the middle layer — precisely the zone where commercial deployment is densest. Its top ranking on the coding agent index shows that Google has made substantive progress on the dimension of "models that plan, use tools, and complete tasks autonomously," not just higher quiz scores. The pricing strategy is even more interesting: Carbon's introductory pricing is about to end, and Neon conveniently becomes the cheaper option. This timing design is clearly intended to steer developers to migrate workloads to Neon. For teams managing API costs, December 1 is a date worth marking.
It's worth noting that launch-week benchmark results are ultimately just early signals. Real coding agent performance depends heavily on the specific codebase, toolchain, and task type. It's advisable to run a round of validation with your own workflows before formally migrating.
One-line Summary
Gemini 4 Neon tops the coding agent index upon release, becomes cheaper than Carbon starting in December, making it Google's most cost-effective frontier coding model.
Paragraph Summary
Google's Gemini 4 family adds Neon, positioned between the speed-focused Hydrogen and the heavyweight Carbon, targeting complex software engineering, knowledge work, and scientific research. Within a day of release, Neon topped the Artificial Analysis Coding Agent Index with 58.6, ranked third on the comprehensive intelligence index with 61, and scored 65.2% on Terminal Bench 4.1. It features a 1 million token context and four thinking levels (defaulting to high), has no free tier, and is available through the paid Gemini API, Gemini for Business, Anti-Gravity, and Pro/Ultra plans. Priced at $3 input / $15 output per million tokens with a 95% discount on cached input, Neon becomes the cheaper frontier coding model after Carbon's price increase on December 1. The article notes that launch-week benchmarks are only early signals and recommends validating with your own workflows before migrating.
Related articles

Gemini 4's Three Coding Models Compared: How to Choose Between Neon, Hydrogen, and Argon
How to choose among Gemini 4's Neon, Hydrogen, and Argon coding models? This article compares official positioning, API pricing and the price-increase date, coding benchmark scores from Artificial Analysis and Cursor, and summarizes developer feedback to offer scenario-based selection advice.

Gemini 4 Neon Hands-On: Crushes Opus at Coding, but the Free Agent Hides Three Major Pitfalls
Gemini 4 Neon tops the Coding Agent Index, beating Carbon and Opus 5.5 in four real code battles—but hidden reasoning tokens bill at $15 and free Tasks silently fails.

Gemini 4 Neon Scores 61 on Intelligence Index: Analyzing a Purpose-Built Model for Long-Running Agents
Gemini 4 Neon scores 61.81% on the Artificial Analysis Intelligence Index, ranking third. We analyze its million-token context, multimodal input, API pricing, and free-tier information gaps, with selection and pilot advice for agent scenarios.