Gemini 4 Neon: #3 on intelligence at 32% lower cost per task
Key takeaways
- Neon scores 61 on the Artificial Analysis Intelligence Index at xhigh and ranks #3 overall, just behind larger and slower models
- At the default high thinking level Google claims a 32% lower cost per task than Carbon and roughly 50% to 80% lower than top competing frontier models
- Security teams get state-of-the-art claims on WebExploitBench v1.2 and SEC-bench Pro, but the cyber variant goes only to trusted defenders via the Fairwind Program
- Most figures come from Google's announcement, so teams should test Neon on their own workloads before switching
Google's Gemini 4 Neon is not the top model on the headline intelligence leaderboard, but it is close, and Google says it costs less per task than Carbon and than competing frontier models. Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, ranking #3 overall, "just behind larger, slower models". The pitch is cost per task, not raw peak capability.
The headline numbers
Google published a long list of results in its 18 November 2026 announcement. The ones that matter most for buyers:
| Measure | Result |
|---|---|
| Artificial Analysis Intelligence Index (xhigh) | 61, #3 overall |
| Vals Index | #2, 70.9% |
| Artificial Analysis Coding Agent Index | #1, 58.6 |
| Terminal-Bench 4.1 | 65.2% |
| FrontierCode 2.0 (default high) | 46.2% |
| SWE-sweep, mini-SWE harness | 12.8% |
| SWE-sweep, Google Antigravity | 39.0% |
| Agent Arena Index | #2, 1336 across 4,200+ sessions |
| Terminal-Bench Science / RiemannBench / HLE-Diamond | 70.5% / 84.0% / 71.3% |
| Input / cached input / output price per million tokens | $3 / $0.15 / $15 |
The context window is 1 million tokens. Neon also takes #1 on Zapier's AutomationBench (56.3%), the Legal Research Bench and the Tax Agent Bench, and posts $16,843 on Andon Labs' Vending-Bench 2, a long-horizon business simulation.
What the data actually shows
The pattern is efficiency. At the default high thinking level, Google says Neon pairs a high Artificial Analysis score with a 32% lower cost per task than its sibling Carbon. It sits on the Pareto frontier (the set of models where no rival is both cheaper and better) on both the Vals Index and Agent Arena, at roughly 50% to 80% lower cost per task than top competing frontier models. At xhigh, where Neon shows its highest score, Google says it still costs 36%+ less per task than competing models at their highest setting.
The most striking number is the harness effect. On SWE-sweep, Neon scores 12.8% on the mini-SWE harness but 39.0% inside Google Antigravity, more than triple. That says as much about the scaffolding around a model as the model itself. A harness is the tooling that feeds a model files, tools and memory; the same weights can look very different depending on it.
The cyber results are the other cluster. Neon reports state-of-the-art results of 76% on the hard tier of AutoBaxBench for secure code generation, 51.26% on WebExploitBench v1.2 at half the cost of competing large models, and 90.5% on SEC-bench Pro at roughly half the cost per challenge. It also reaches 70.7% on CWE-bench v2, a 3.4% uplift over Carbon. On ExploitGym it matches frontier performance with half of Carbon's tokens.
What the data does not tell us
Most of these figures are Google's own presentation of third-party benchmarks, and the announcement does not disclose test conditions for every claim. Several are "state-of-the-art" or "#1" results on newer benchmarks, such as Agent Arena Index, which Google itself calls new. The lowest fabrication rate is from internal Google Antigravity testing, so it cannot be independently compared.
Cost per task is also workload dependent. The 32% saving against Carbon and the 50% to 80% saving against rivals are Google's figures at specified thinking levels; a team that needs xhigh reasoning on every call will see a different bill from one running at high. The headline per-token prices say little without knowing how many tokens a task burns. Cached input at $0.15 per million, a 95% discount, helps only where prompts repeat.
Neon is also #3, not #1, on the main intelligence index. Larger, slower models still lead there.
How this compares
The cyber story echoes the gated approach described in Gemini 4 Argon is Google's strongest model, and only security teams get it: Neon for cyber is rolling out to trusted defenders in the coming days through the same Fairwind Program. Security teams weighing options can also read Gemini 4 Carbon vs Argon: which model security teams should pick.
So what?
For UK and US enterprise teams, the practical question is cost per completed task, not list price. Availability through the Gemini API, Gemini for business and, for paid Pro and Ultra subscribers, the Gemini app and Antigravity means it can be piloted now without a special agreement.
Security teams should note the safety stance. Google says Neon keeps Carbon-level protections against severe misuse while refusing just 1.98% of tasks on Terminal-Bench 4.0 and 0% on Terminal-Bench-Science, meaning fewer blocked legitimate requests. Whether that holds for a team's own penetration-testing work is something only a trial will show.
Sensible steps: run a representative sample of real tasks at both high and xhigh, record tokens per completed task rather than per-token cost, and compare against the model already in use. The benchmark lead matters less than whether Neon's 32% saving survives contact with a team's own workload.
Frequently asked questions
How much does Gemini 4 Neon cost?
Google lists $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens, with a 1 million-token context window.
Who can use Gemini 4 Neon today?
It is rolling out in the Gemini API in Google AI Studio, Gemini for business, Google Antigravity for paid AI Pro and Ultra subscribers, and the Gemini app for Pro and Ultra subscribers.