DEV Community

Lâm Phú linh
Lâm Phú linh

Posted on

Gemini 4 Neon: Google's Middle Child Is Now Its Best Agent — #1 on the Coding Agent Index, 32% Cheaper Than Carbon

Google announced Gemini 4 Neon on November 18, 2026 — the third Gemini 4 model in seven weeks, and the one that finally makes the family make sense. Carbon is the heavyweight for your hardest, long-horizon projects. Hydrogen is the fast everyday runner. Neon sits between them, and the pitch is simple: frontier-level reasoning for the complex, high-frequency work you actually run every day, at a lower cost per task.

The headline numbers

Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, ranking #3 overall — just behind larger, slower models. On the Vals Index it ranks #2 at 70.9% and lands on the Pareto frontier. And on Arena's brand-new Agent Arena Index, built from 4,200+ real-world agent sessions, Neon debuts at #2 globally with a score of 1336 — the best-ranked Google DeepMind model on that leaderboard, ahead of Carbon (1303), Hydrogen (1271), and Argon (1229).

It also ranks #2 on steerability — how cleanly it adapts to mid-task user corrections — which matters more than it sounds. An agent that can't take a correction mid-flight is an agent you have to restart.

Coding is where Neon leads outright

Neon ranks #1 overall on the Artificial Analysis Coding Agent Index at 58.6, with 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 more than triples to 39.0% inside Google Antigravity.

The engineering pitch is about long sessions, not single prompts: Neon is built to sustain performance through complex, tool-heavy tasks end to end. It cuts down on redundant file reads, shares progress as it works, and handles context compaction reliably in long sessions. In Antigravity it can act as a fast, high-judgment orchestrator directing parallel Hydrogen subagents. For production-ready code, Google says it handles multi-file debugging and refactoring across repositories with targeted, surgical edits.

Knowledge work and science

Beyond code, Neon ranks #1 on both Vals' Legal Research Bench and Tax Agent Bench, #1 on most Surge AI benchmarks across finance, accounting, spreadsheet modeling, and complex constraint following, and #1 overall on Zapier's AutomationBench (56.3%) and Andon Labs' Vending-Bench 2 ($16,843) — a long-horizon simulation of running a business. It scores 73.5% on Chartography for reading dense professional charts and 75.8% on Agents' Last Exam v1.1, on par with leading models.

On scientific and mathematical reasoning it is best in class on Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%).

The cost math

Neon is priced at $3 per million input tokens ($0.15 per million cached input tokens, a 95% discount) and $15 per million output tokens, with a 1 million-token context window. At its default high thinking level, Google reports a 32% lower cost per task than Carbon, and roughly 50% to 80% lower cost per task than top competing frontier models. Even at xhigh — where Neon posts its highest intelligence score — it still costs 36%+ less per task than competing models at their highest setting.

Four thinking levels scale steadily, so teams can dial intelligence and token spend to what each workload needs instead of paying frontier prices for routine work.

Built for cyber defense, not just code

There's a dedicated variant, Gemini 4 Neon for cyber, rolling out to trusted defenders through the Fairwind Program. It produces more secure code than Carbon across easy, medium, and hard challenges on AutoBaxBench, reaching 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 hits 90.5% while costing roughly half as much per challenge, and on CWE-bench v2 it reaches 70.7%, a 3.4% uplift over Carbon. On ExploitGym it achieves frontier-level performance using half the tokens of Carbon.

The behavior numbers are the interesting part

Benchmarks tell you what a model can do. The behavior numbers tell you what it will actually do in a long session — and this is where Google is making its case:

  • Its lowest fabrication rate yet in internal Google Antigravity testing.
  • 0.0% cheating on Terminal-Bench 4.0, and shortcuts on just 2.45% of tasks on Google's coding cheating evaluation — the lowest rate of any Gemini 4 model.
  • A 0.09% misaligned action rate on the agentic safety suite, and a 99.2% pass rate on the agent honesty evaluation.
  • Under-service reduced by 43%–56% and overreach by 15%–35% versus Carbon, across effort levels.
  • On Gray Swan's indirect prompt injection benchmark, Neon resisted every attack at its highest thinking level — a 0.0% attack success rate — while refusing just 1.98% of tasks on Terminal-Bench 4.0 and 0% on Terminal-Bench Science.

In other words: it finishes the work, doesn't invent results, doesn't cut corners, and doesn't overstep its scope. For agents that run unattended for hours, that behavior profile matters more than another point on a benchmark.

Where to get it

Neon is rolling out in the Gemini API in Google AI Studio, in Gemini for business (via API and model dropdown), for paid Google AI Pro and Ultra subscribers in Google Antigravity, and for Pro and Ultra subscribers in the Gemini app. Neon for cyber reaches eligible Fairwind Program members over the coming days.

The takeaway

Google's message with Neon is that the frontier is no longer the point — the middle is. The most expensive model in the family isn't the one that ships the most work; the one that pairs near-frontier intelligence with fast, cheap, well-behaved execution is. If your agents spend their days on complex coding, analysis, and multi-step workflows rather than once-a-quarter moonshots, Neon is the tier Google built for you — and the benchmarks say it's the tier where Google is now winning.

Source: Google, "Introducing Gemini 4 Neon" (November 18, 2026).

Top comments (0)