Trending · From the catalog
Gemini 4 Neon: The Mid-Tier Model That Just Broke the Speed-Cost Curve
Google's Gemini 4 Neon hits #2 on the Vals Index at a fraction of the cost of frontier models. We break down the benchmarks, the pricing math, and why this changes how you choose models for your AI coding agents.
The AI coding community spent the last week arguing about which model to use for what, a debate that usually ends in a shrug and a roulette spin. Then Google dropped Gemini 4 Neon, a mid-tier model that scored 61 on the Artificial Analysis Intelligence Index and ranked #2 on the Vals Index. It is not the biggest model, but it is fast, cheap, and surprisingly good at long-horizon tasks.
Neon is priced at $3 per million input tokens and $15 per million output tokens, with a 1 million-token context window. It sits on the Pareto frontier, meaning you get near-frontier intelligence for 32% to 80% less per task than the top competing models. For developers running agents that burn through tokens on multi-step coding tasks, that is the headline.
The Naming Chaos Is Real
Google has released four Gemini models this month, and the naming convention is now a full-time job. One commenter on Hacker News pointed out that you might as well use a roulette wheel to pick which model to use for a given task. The lineup now includes Carbon, Argon, Astra, Fable, Neon, Sol, Opus, Hydrogen, Terra, Sonnet, and a rumored Helix.
Neon is the long-awaited successor to 3.1 Pro, and it lands in the 'Large' class alongside Sol and Opus. It is not the flagship, but it is close enough that the distinction matters less than the price tag. If you are trying to figure out which model to use for what, the answer is increasingly: it depends on your budget, not just your task.
- Neon ranks #2 on the Vals Index with a score of 70.9%.
- It lands on the Pareto frontier, meaning best intelligence-to-cost ratio.
- The naming scheme is now so confusing that developers are asking for a model selection guide.
The Benchmarks That Matter
Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, ranking #3 overall. It is not the top model, but it is close. On the Vals Index, it ranks #2 at 70.9%, just behind Fable. On the Legal Research Bench and Tax Agent Bench, it ranks #1. On Surge AI, it ranks #1 on most benchmarks across finance, accounting, spreadsheet modeling, and complex constraint following.
For coding specifically, Neon ranks #1 overall on the Artificial Analysis Coding Agent Index with a score of 58.6. It scores 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default high setting. On SWE-sweep, it leads the mini-SWE harness at 12.8% and jumps to 39.0% in Google Antigravity, a state-of-the-art result that shows what the Antigravity harness unlocks.
- Neon scores 61 on the Artificial Analysis Intelligence Index at xhigh thinking.
- It ranks #1 on the Artificial Analysis Coding Agent Index at 58.6.
- On SWE-sweep, it reaches 39.0% in Google Antigravity, a state-of-the-art result.
- It scores 73.5% on Chartography for reading dense professional charts.
The Pricing Math Is the Real Story
Neon is priced at $3 per million input tokens, with a 95% discount on cached input tokens at $0.15 per million. Output tokens cost $15 per million. At its default high thinking level, it delivers 32% lower cost per task than Carbon. At xhigh thinking, it costs 36% or more less per task than competing models at their highest setting.
That is the number that matters for developers running agents. If you are using a model like Fable or Carbon for long-running coding tasks, you are paying a premium for marginal gains. Neon sits on the Pareto frontier on both the Vals Index and Agent Arena at high, with roughly 50% to 80% lower cost per task than top competing frontier models. For a team running 100 agents a day, that is a meaningful difference.
- Neon costs $3 per million input tokens and $15 per million output tokens.
- Cached input tokens are $0.15 per million, a 95% discount.
- At high thinking, Neon costs 32% less per task than Carbon.
- At xhigh thinking, it costs 36% or more less per task than competing models.
The Cybersecurity Angle Nobody Talked About
Neon produces more secure code than Carbon across all easy, medium, and hard challenges on AutoBaxBench, achieving state-of-the-art on the hard tier at 76%. On WebExploitBench v1.2, it is state-of-the-art at 51.26%, at half the cost of competing large models. On SEC-bench Pro, it achieves 90.5%, costing roughly half as much per challenge.
For developers building AI coding agents that need to generate secure code from the start, this is a significant advantage. Neon for cyber is rolling out to trusted defenders as part of the Fairwind Program. If your agent is writing code that will run in production, the security benchmarks matter as much as the coding benchmarks.
- Neon achieves state-of-the-art on AutoBaxBench hard tier at 76%.
- It scores 51.26% on WebExploitBench v1.2 at half the cost of competing models.
- It achieves 90.5% on SEC-bench Pro at roughly half the cost per challenge.
- Neon for cyber is rolling out to trusted defenders via the Fairwind Program.
What This Means for Your AI Coding Stack
Neon is not the biggest model, but it is the one that makes the most sense for most teams. It is fast, cheap, and good enough for the vast majority of coding tasks. If you are running agents that burn through tokens on multi-step tasks, Neon is the model to use. If you need the absolute top of the leaderboard for a specific task, use Fable or Carbon. But for the 80% of tasks that do not need the top 1%, Neon is the right choice.
The real question is not whether Neon is good enough. It is. The question is whether you are overpaying for models that are only marginally better. Neon changes the math. It is the mid-tier model that finally makes the speed-cost curve work for developers.
- Use Neon for the 80% of tasks that do not need the top 1% of model performance.
- Use Fable or Carbon for tasks that require the absolute top of the leaderboard.
- Neon is available in the Gemini API, Google AI Studio, Gemini for business, and Google Antigravity.
- It is also available for Pro and Ultra subscribers in the Gemini app.
Skills to try
- gemini-api-cli: Call the Gemini API from the shell with gemini-api. Use for media generation (images, TTS, music, video), file and YouTube analysis, audio/video transcription, Deep Research, and s…
- agent-platform-architecture: Designs, engineers, and validates enterprise Gemini platform architectures. Use when designing multi-project hierarchies (Host, Central Governance Hub, Runtime Spoke, Tool/MCP Spok…
- code-review: Thorough code review — checks correctness, security, performance, readability, and test coverage. Gives actionable feedback ranked by severity.
- senior-frontend: Build and review React, Next.js, TypeScript or Tailwind frontends: components, scaffolding, performance and bundle size. Use whenever the user wants a new component or app, a page …
- web-game: Scaffolds a browser-based game with Phaser 3, PixiJS, or Three.js including game loop, asset pipeline, responsive canvas, input handling, audio, save/load, and deployment to itch.i…
Sources
Written by
Skills-Hub Team
Skills-Hub is the open registry for AI coding skills, with SKILL.md files synced daily from Anthropic, Google, Microsoft, and 90+ official sources. Free + MIT.