Summary: Google launched Gemini 4 Neon on November 18, 2026 — a mid-tier model slotted between Carbon and Hydrogen in the Gemini 4 family, aimed at complex coding, analytical, and enterprise work. It is live today in the paid Gemini API at a stated price, with strong safety scores attached. The catch for builders: nearly every performance number in the announcement was measured by Google or inside Google’s own tooling.
What Neon is and what it costs
Google positions Neon as the family’s balance point. Carbon handles the hardest long-horizon projects, Hydrogen does fast everyday execution, and Neon takes high-frequency workflows that still need frontier-level reasoning. In Google Antigravity it can also run as an Orchestrator directing parallel Hydrogen subagents; Google says it cuts redundant file reads and handles context compaction reliably across long sessions.
The pricing is explicit: $3 per million input tokens, $15 per million output tokens, and a 1 million-token context window. Cached input costs $0.15 per million tokens — a 95% discount, which matters a lot if your agents replay long contexts every turn. Four thinking levels let you trade token spend against reasoning depth.
Same model, two very different coding scores
As reported in Google’s announcement, Neon ranks #3 overall with a score of 61 on the Artificial Analysis Intelligence Index at its xhigh setting, #2 on the Vals Index (70.9%), and #2 globally on Arena’s new Agent Arena Index (1336, across 4,200+ real-world agent sessions). On coding it claims #1 on the Artificial Analysis Coding Agent Index (58.6), 65.2% on Terminal-Bench 4.1, and 46.2% on FrontierCode 2.0 at its default high setting.
The number that should slow you down is on SWE-sweep: Neon scores 12.8% on the mini-SWE harness and more than triples to 39.0% inside Google Antigravity. Harness choice alone moves the result by a factor of three, and the higher figure comes from Google’s own harness. Until something independent reproduces it, treat the 39% as an Antigravity number, not a model number.
On cost per task — Google’s own measurement again — Neon at its default high setting runs 32% cheaper than Carbon, and sits 50% to 80% below top competing frontier models at high on the Vals and Agent Arena Pareto frontiers. At xhigh it still costs 36%+ less per task than competitors at their highest setting. Related: Gemini 4 Carbon ships at Argon pricing: judge it per task.
Safety scores, with their conditions attached
The safety results are the announcement’s most concrete claims. On Gray Swan’s indirect prompt injection benchmark, Neon posted a 0.0% attack success rate — but only with real-time safeguards enabled at its highest thinking level, so that is a configuration you must actually turn on, not a default. Google also reports a 99.2% pass rate on its agent honesty evaluation, a 2.45% shortcut rate on its coding cheating evaluation (the lowest of any Gemini 4 model), and a 43%–56% reduction in under-service and 15%–35% reduction in overreach versus Carbon’s baseline across effort levels.
A separate “Neon for cyber” variant posts strong numbers — state of the art on AutoBaxBench’s hard tier (76%) and 51.26% on WebExploitBench v1.2 at half the cost of competing large models — but it is not generally available. It reaches trusted defenders in the coming days through the Fairwind Program.
Our take: Run one test this week; don’t migrate anything yet. Neon clears the bar for a cheap, reversible experiment: it is in the paid API today at a stated price, with a 1M-token context and cached input at $0.15 per million tokens. But the case for switching rests on numbers Google measured, and the 12.8%-versus-39.0% SWE-sweep gap shows that the harness alone can triple a coding score — so assume, until shown otherwise, that your stack looks more like the generic harness than Antigravity. Pick five tasks from your real backlog, run Neon at its default high setting, and compare finished-task quality, total cost including retries, and latency against your current model. If independent reruns outside Google’s harness confirm the coding-agent and cost-per-task numbers, this becomes an adoption decision rather than a test. Teams whose workloads fit a smaller, cheaper model can skip this entirely, and anyone who needs the cyber variant today is out of luck — it is gated behind the Fairwind Program.
Where you can use it today
No waitlist for the base model: it is rolling out now in the Gemini API paid tier in Google AI Studio and Android Studio, in Gemini for business via API and model dropdown, in Google Antigravity for paid Google AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra. That availability is what makes the afternoon experiment possible — you are judging a live product, not a roadmap. Google’s post continues into enterprise rollout detail beyond what this article covers; the pricing, benchmarks, and safety results above are what its text shows.
Sources
View the original post on X
