Google Launches Gemini 4 Neon, Beating Carbon on Coding Agents at Lower Cost per Task
On this page 9 sections
TL;DR
Gemini 4 Neon shipped Nov 18, Google's fourth Gemini 4 model in about seven weeks, and the middle tier undercuts the flagship on cost-adjusted coding work. Google's launch post ranks it #1 on the Artificial Analysis Coding Agent Index at 32% lower cost per task than Gemini 4 Carbon, for $3/$15 per million tokens over a 1M-token window. One more number should change how you read agent benchmarks: the same weights score 12.8% on SWE-sweep alone and 39% inside Antigravity.
Four elements in seven weeks
Google names its tiers after elements in an order the periodic table does not recognize: Carbon (atomic number 6) is the flagship, Neon (10) the mid tier, Hydrogen (1) the budget one. The cadence is the louder fact. Arena counted 28 days between Argon and Carbon, 15 between Carbon and Hydrogen, and 6 between Hydrogen and Neon. Latent Space reports the six-week-old Argon is already retired, and the Pro and Flash names from Gemini 1 through 3 are gone.
Neon is rolling out through the paid Gemini API in AI Studio and Android Studio, Gemini for business, Antigravity for paid AI Pro and Ultra subscribers, and the model dropdown in the Gemini app for the same tiers. Cursor, GitHub Copilot across its app, CLI and VS Code, and LangChain's managed agents all shipped support on day one.
The coding numbers
Google's table puts Neon first 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. Independent of Google, Cursor measured 63.7% on CursorBench 4.1 at default effort, against 57.3% for Hydrogen.
Artificial Analysis adds the cost side: at high effort Neon matched Carbon's 60 on the Intelligence Index for $3.74 per task, and xhigh nudges it to 61, third overall behind larger, slower models. Treat the first-place rankings as Google-reported until the boards refresh; the independent indexes below mostly agree, with two specific exceptions.
The harness tripled the score
On SWE-sweep, Neon leads the mini-SWE harness at 12.8%. The same model inside Antigravity scores 39%, which Google calls state of the art. Nothing about the weights changed between those two runs.
What Antigravity adds is structure: Neon acts as a fast, high-judgment orchestrator that directs parallel Hydrogen subagents in isolated workspaces, cuts redundant file reads, shares progress as it works, and handles context compaction over long sessions. It is the same engine with a much better transmission; the harness decides how much of the model's power reaches the road. The New Stack read the jump the same way.
The practical version: before you switch your default model, run your own eval inside your own harness. A model that wins in Google's harness can behave differently in yours, in either direction.
The independent boards
On the Vals Index, Neon debuts second at 70.9%, about 0.4 points behind Claude Fable 5.2 at less than half the cost per task, and it takes first on Vals' Legal Research Bench and Tax Agent Bench. On Agent Arena it debuts second with 1,336 points across about 4,200 real sessions, a $2.82 median task cost, and second place on steerability, which measures how cleanly it absorbs a mid-task correction. Carbon still leads the separate Text Arena.
Two results cut the other way. On Vals' Code Migration suite, Neon averaged 160 turns and scored below both Carbon and Hydrogen, which needed 234 and 574 turns. On GDP.pdf at Artificial Analysis, high effort beat xhigh, 33.4% to 28.0%, so the extra thinking sometimes subtracts. One early user's logs from 13 Antigravity sessions show a 93% cache hit rate and zero shell commands when the prompt said to avoid the terminal, alongside 57 capacity errors and four stalled runs, normal wobble for a rollout this fast.
Elsewhere in Google's table: 70.5% on Terminal-Bench Science, 84% on RiemannBench, 71.3% on HLE-Diamond, first on Zapier's AutomationBench at 56.3%, and first on Vending-Bench 2 with $16,843 in a long-horizon simulation of running a vending business. The vending machine is now an AI benchmark.
The rate card
Neon costs $3 per million input tokens, $15 per million output, and $0.15 cached, a 95% discount, over a 1M-token window with up to 1M output tokens. Per token that is the priciest slot in the family: Hydrogen runs $1.50/$7.50, and Carbon and Argon sit at introductory $2/$10 until they double to $4/$20 on Dec 1, per The Decoder.
Per task the order flips. Google's chart has Neon at 32% lower cost per task than Carbon at the default high setting, and 50 to 80% below top frontier models on the Vals and Agent Arena cost frontier; at xhigh it still comes in over 36% cheaper per task than rivals at their maximum settings.
If your agent loop rereads the same files all day, the $0.15 cached line matters more than the sticker prices. The first read pays full fare and every reread is pocket change.
The defender tier
The launch includes a cyber variant, gated to vetted defenders through the Fairwind Program over the coming days. The numbers: 51.3% on WebExploitBench v1.2, state of the art at half the cost of comparable large models; 90.5% on SEC-bench Pro, also at roughly half the cost per challenge; 70.7% on CWE-bench v2, 3.4 points over Carbon; and state of the art on the hard tier of AutoBaxBench at 76%.
The gating makes sense: a model this good at finding and validating exploitable bugs is the last model you hand to everyone with an API key.
The fine print
The behavior numbers are the best in the family so far. Neon took shortcuts on 2.45% of Google's coding cheating evaluation, the lowest rate of any Gemini 4 model, cheated on zero Terminal-Bench 4.0 tasks, passed 99.2% of the agent honesty evaluation, and posted a 0.09% misaligned action rate on the agentic safety suite. Under-service dropped 43 to 56% and overreach 15 to 35% against Carbon's baseline.
On Gray Swan's indirect prompt injection benchmark, Neon with real-time safeguards at its highest thinking level resisted every planted attack, a 0.0% success rate only Gemini models have reached so far. It stays usable while doing it: 1.98% refusals on Terminal-Bench 4.0 and zero on Terminal-Bench Science, low for a model carrying Carbon-grade cyber safeguards.
For a sense of what it does with long context, Google's demos include turning a single image into a 100,000-cell Google Sheet driven by more than 500,000 live formulas, with no embedded images and no Apps Script.
Key Takeaways
- Gemini 4 Neon shipped Nov 18, the fourth Gemini 4 model in about seven weeks, after gaps of 28, 15 and 6 days; it sits between Carbon and Hydrogen, and Argon is already retired.
- Priced $3/$15 per million tokens with $0.15 cached input and a 1M-token window; per task Google puts it 32% below Carbon at the default high setting and 50 to 80% below frontier rivals on the Vals and Agent Arena cost frontier.
- Coding results: #1 on the AA Coding Agent Index at 58.6 and 65.2% on Terminal-Bench 4.1 per Google; 63.7% on CursorBench 4.1 per Cursor, against Hydrogen's 57.3%.
- The harness effect is the headline for builders: 12.8% on SWE-sweep solo, 39% inside Antigravity, same weights. Test and report agents as model plus harness.
- Independent boards mostly agree: second on Vals at 70.9%, 0.4 behind Claude Fable 5.2 at under half the cost per task, second on Agent Arena. Watch Code Migration and GDP.pdf, where harder thinking scored worse.
- A cyber variant leads WebExploitBench v1.2 at 51.3% and SEC-bench Pro at 90.5%, gated to Fairwind defenders because the same skill that patches bugs also finds them.
Sources: Google, Introducing Gemini 4 Neon, Nov 18, 2026, Gemini 4 Neon model card, The Decoder, Nov 18, 2026, Latent Space AI News, Nov 18, 2026, The New Stack, Nov 19, 2026, Cursor on X, Vals on X, Arena on X