Google’s most interesting claim about Gemini 4 Neon isn’t a leaderboard rank. It’s that the model matched or beat Gemini 4 Carbon on overall quality at three times the speed and 60% of the cost, at least in an early evaluation from Glean. If that holds up outside vendor-picked testimonials, the case for paying for the biggest model in any lineup gets a lot weaker.
Google announced Neon on Nov 18, 2026, pitching it as frontier-level intelligence with fast end-to-end task completion for software engineering, knowledge work, and science.
What the benchmarks say
Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level. That puts it #3 overall, just behind larger and slower models. On the Vals Index it ranks #2 and sits on the Pareto frontier, and it also debuts #2 on Arena’s new Agent Arena Index with a 1336 score across more than 4,200 agent sessions.
The coding numbers are the strongest part of the pitch. Neon ranks #1 on the Artificial Analysis Coding Agent Index (58.6), hits 65.2% on Terminal-Bench 4.1, and 46.2% on FrontierCode 2.0. On SWE-sweep it scores 12.8% on the mini-SWE harness, but jumps to 39.0% inside Google Antigravity. That gap is telling. Google is saying the harness matters as much as the model, which is the same argument Anthropic makes with Claude Code and OpenAI with Codex.
On the science side, Google claims best-in-class results on Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%). Treat those with the usual caution. Most of these benchmarks are new, and several are run by third parties with limited track records.
Built for agents, not just chat
Google frames Neon as a long-running agent model. It cuts redundant file reads, reports progress as it works, and handles context compaction in long sessions. It can also act as an orchestrator for parallel Hydrogen subagents in Antigravity. That’s a clear nod to the multi-agent workflows developers are already building around Claude and GPT models.
Steerability is the quieter but more useful claim. Neon ranks #2 on that measure, meaning it adapts cleanly when you correct it mid-task. Google also says it recorded its lowest fabrication rate yet, though that figure comes from internal Antigravity testing and can’t be checked independently.
The enterprise results are broad. Neon is #1 on Legal Research Bench and Tax Agent Bench, #1 on Zapier’s AutomationBench (56.3%), and #1 on Andon Labs’ Vending-Bench 2 ($16,843). It handles spreadsheets, PDFs, charts, audio, and video in one 1 million-token window.
Pricing and availability
Neon costs $3 per million input tokens and $15 per million output tokens. Cached input is $0.15 per million, a 95% discount. That input and output pricing matches Claude Sonnet’s list price, so the real fight is on cost per task. Google says Neon is 32% cheaper per task than Carbon at its default high setting, and 50% to 80% cheaper than top competitors on the Pareto charts.
It’s available now in the Gemini API through Google AI Studio, in Gemini for business, in Google Antigravity for paid AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra users. Four thinking levels let teams trade intelligence against token spend per workload.
Does it matter?
Yes, with caveats. The benchmark spread is wide, the price is competitive, and the cost-per-task argument speaks to what teams actually pay for. But the testimonials are curated, and the biggest coding gains depend on Antigravity, which locks you further into Google’s tooling. If you run agents in production, test Neon against your own workloads at the high setting before switching. The headline ranks are good. Your own evals are better.



