Skip to content
ISSUE 001·LIVE·13:04 IL
← Journal/2026-11-18·9 min·google

Gemini 4 Neon: Google's fast reasoning model for production coding agents

Google launched Gemini 4 Neon on November 18, 2026: a fast reasoning model positioned between Carbon and Hydrogen, priced at $3 per million input tokens with a 1M context window. Google claims #1 on the Artificial Analysis Coding Agent Index, state-of-the-art SWE-sweep results inside Antigravity, and 32% lower cost per task than Carbon. Here is what shipped, what the benchmarks say, and what to watch.

By Harel Asaf·AI Builder·Tel Aviv

Google released Gemini 4 Neon today, and it is the most interesting Gemini 4 variant for anyone running coding agents in production: a fast reasoning model with a 1 million-token context window, priced at $3 per million input tokens ($0.15 cached) and $15 per million output tokens, available now in the Gemini API, AI Studio, Antigravity, and the Gemini app for Pro and Ultra subscribers. Google's pitch is simple: near-Carbon quality at roughly 3x the speed and 60% of the cost. All benchmark numbers below come from Google's announcement post unless noted otherwise, so treat them as vendor claims until third-party evals land.

Where Neon sits in the Gemini 4 family

Google now ships a three-tier lineup, and the framing matters for how you route work. Carbon is the biggest, slowest model for the hardest long-horizon projects. Hydrogen handles fast everyday execution. Neon sits between them: frontier-level reasoning for complex, high-frequency workflows where Carbon is too slow and too expensive to run all day. Inside Google Antigravity, Google explicitly positions Neon as a fast orchestrator that can direct parallel Hydrogen subagents, which tells you how they expect agent hierarchies to be built: Neon plans and judges, Hydrogen executes the cheap steps.

The benchmark numbers that matter

On general intelligence, Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, ranking #3 overall behind larger, slower models, and #2 on the Vals Index at 70.9% where it sits on the Pareto frontier. For agentic work, it debuts at #2 on Arena's new Agent Arena Index (1336 across 4,200+ real-world agent sessions) and ranks #2 on steerability, meaning it adapts cleanly when you correct it mid-task.

For coding specifically, Google claims #1 overall on the Artificial Analysis Coding Agent Index (58.6), 65.2% on Terminal-Bench 4.1, and 46.2% on FrontierCode 2.0 at the default high setting. The number with the biggest asterisk: on SWE-sweep Neon leads the mini-SWE harness at 12.8% but jumps to a state-of-the-art 39.0% when run inside Google Antigravity. That gap is the harness, not the weights, and Google says so themselves by presenting it as what Antigravity unlocks. Expect the 12.8% figure to be the one that transfers to your own scaffolding.

On science and math, Neon posts best-in-class Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%), plus #1 on Legal Research Bench, Tax Agent Bench, Zapier's AutomationBench (56.3%), and Andon Labs' Vending-Bench 2.

Cost per task is the real argument

The pricing story is where Neon gets interesting for production. At the default high thinking level, Google claims a 32% lower cost per task than Carbon while keeping a high intelligence score, and roughly 50% to 80% lower cost per task than top competing frontier models on Vals and Agent Arena. The xhigh level adds deeper multi-step reasoning for the hardest workloads at what Google says is still 36%+ less per task than competitors at their highest settings. Neon has four thinking levels, so you can dial intelligence and token spend per workload instead of per deployment. If those task-level numbers hold under third-party measurement, Neon becomes the default middle tier for agent loops that run thousands of times a day.

The cyber push

Google is making a targeted play for security teams: a Gemini 4 Neon for cyber variant rolls out to trusted defenders through the Fairwind Program over the coming days. Claimed results include state-of-the-art on AutoBaxBench hard tier (76%) for generating secure code, SOTA on WebExploitBench v1.2 (51.26%) at half the cost of competing large models, SOTA on SEC-bench Pro vulnerability finding (90.5%), and 70.7% on Collinear's CWE-bench v2 for patching, a 3.4% uplift over Carbon. Access is gated: you request it through a Cloud account team.

Safety and behavior claims worth noting

Two claims stand out for agent builders. First, Google says Neon refuses far less legitimate work than typical frontier models in sensitive domains: 1.98% refusal on Terminal-Bench 4.0 and 0% on Terminal-Bench Science, while keeping Carbon-level misuse safeguards. Second, on behavior discipline: 0.0% cheating on Terminal-Bench 4.0, shortcuts on 2.45% of tasks (the lowest of any Gemini 4 model), a 0.09% misaligned action rate on Google's agentic safety suite, and a 99.2% pass rate on its agent honesty evaluation. On Gray Swan's indirect prompt injection benchmark, Google says Neon with real-time safeguards at its highest thinking level resisted every attack, a 0.0% attack success rate they say only Gemini models have reached. These are internal evaluations; the model card has the details.

What I could not verify

Everything above originates from Google's launch post and partner quotes (Glean and others praising code exploration and ML engineering). I did not run the model. Third-party leaderboards like Artificial Analysis publish their own measurements over the coming days, and those are the numbers to wait for before rerouting production traffic. The Antigravity-dependent SWE-sweep result, the cost-per-task curves, and the refusal-rate claims are the three places where independent verification will matter most.

Frequently asked questions

How much does Gemini 4 Neon cost?

$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. Google claims cost per task runs 32% below Gemini 4 Carbon at the default high thinking level.

Where can I use Gemini 4 Neon today?

In the Gemini API paid tier via Google AI Studio and Android Studio, in Gemini for business (API or model dropdown), in Google Antigravity for paid Google AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. The cyber variant requires Fairwind Program access.

How is Neon different from Gemini 4 Carbon and Hydrogen?

Carbon is the largest model for the hardest long-horizon work, Hydrogen is the fast everyday execution model, and Neon balances depth and efficiency for complex, high-frequency workflows. Google positions Neon as an orchestrator that can direct parallel Hydrogen subagents inside Antigravity.

Is Gemini 4 Neon good for coding agents?

On Google's numbers, yes: #1 on the Artificial Analysis Coding Agent Index, 65.2% on Terminal-Bench 4.1, and strong long-session behavior with reliable context compaction. But its headline SWE-sweep result (39.0%) depends on the Antigravity harness; on the mini-SWE harness it scores 12.8%, so validate against your own scaffolding.

What thinking levels does Neon support?

Four levels, scaling from fast and cheap up to xhigh for the hardest engineering, analytical, and scientific workloads. The top intelligence scores (like the 61 on the Artificial Analysis Intelligence Index) come from the xhigh setting.

Is Gemini 4 Neon open weights?

No. It is a proprietary model available through Google's API and products. There is no indication of a weights release.

Should I switch my production agents to Neon?

Wait for third-party evals if your workload is sensitive to the vendor-claimed numbers, especially cost per task. If you are already on Gemini 4 Carbon for high-frequency agent loops, Neon is designed for exactly your case and is worth benchmarking on your own tasks this week.

Sources

Related reading

  • GPT-6.2 Astra and Sol: OpenAI's answer to the price-performance war

    OpenAI released GPT-6.2 Astra and Sol on November 17, 2026: a frontier model for the hardest professional work at $8 per million input tokens, and an everyday model holding GPT-6.1 Sol's prices. The launch leans on cost-per-task claims across legal, science, coding and 3D benchmarks, ships both models in ChatGPT Work, Codex and the API, and includes a rare disclosure: Astra is OpenAI's first external model to hit the High threshold for AI self-improvement.

  • Microsoft's seven MAI models: the price-performance play for 24/7 coding agents

    At Ignite 2026, Microsoft AI released seven models led by MAI Code 2, which it claims reaches 98% of GPT-6.1 Sol's coding quality at 40% of the price, plus a cyber model at under 2% of a frontier competitor's cost per task, a routing model, and Licensed Weights you run in your own Azure tenant. The strategy is clear: not benchmark trophies, but agents cheap enough to run around the clock.

  • Qwen4-Max: Alibaba's 2.4T flagship API model claims 62% lower cost per task than its predecessor

    Alibaba's Qwen team released Qwen4-Max, the first model in the Qwen4 family: a 2.4-trillion-parameter sparse MoE with 71B active parameters, a 1M-token context window, thinking modes with tunable effort, and both OpenAI-compatible and Anthropic-compatible endpoints. Vendor-reported numbers show up to 49% fewer tokens on agentic coding tasks and a 62% drop in average cost per task versus Qwen3.8-Max. Input starts at 1.5 dollars per million tokens.

The build log

Follow the work.

New essays and prototypes, in your feed reader. Email subscriptions are paused while I connect a working delivery service. No email is collected here.

Follow via RSS