Google introduced Gemini 4 Neon on Wednesday, the latest addition to its Gemini 4 model family and, as The Decoder counted it, the fourth new Gemini model Google has shipped in a month. Google positions Neon as a mid-tier workhorse that pairs frontier-level intelligence with fast end-to-end task completion across software engineering, knowledge work, and scientific workflows — and prices it well below the larger models it trails on raw capability.
"Today, we're introducing Gemini 4 Neon — bringing frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows," Google DeepMind's SVP and Chief AI Architect wrote in the announcement. SiliconANGLE reported that the mid-tier model runs three times faster than its larger sibling Carbon while winning on some metrics. For more context on this story, see our ongoing AI industry coverage.
Where Neon ranks on the benchmarks
On the Artificial Analysis Intelligence Index, Neon scores 61 at its xhigh thinking level, ranking #3 overall — just behind larger, slower models, Google said. On the Vals Index it ranks #2 at 70.9 percent and sits on the Pareto frontier of cost versus performance. It takes the #1 spot on both the Legal Research Bench and Tax Agent Bench, and Google cited Surge AI evaluations ranking it #1 on most benchmarks covering finance, accounting, spreadsheet modeling, and complex constraint following.
The agent-flavored scores are equally prominent in the announcement: #1 overall on Zapier's AutomationBench at 56.3 percent, #1 on Andon Labs' Vending-Bench 2 long-horizon business simulation with $16,843, 75.8 percent on Agents' Last Exam v1.1, and 73.5 percent on Chartography for reading dense professional charts. On Arena's new Agent Arena Index, built from more than 4,200 real-world agent sessions, Neon debuts #2 globally with a score of 1336. In scientific and mathematical reasoning, Google reports best-in-class results on Terminal-Bench Science (70.5 percent), RiemannBench (84.0 percent), and HLE-Diamond (71.3 percent).
Built for long-running coding agents
Coding is where Google is aiming Neon hardest. The model ranks #1 overall on the Artificial Analysis Coding Agent Index at 58.6, posts 65.2 percent on Terminal-Bench 4.1, and scores 46.2 percent on FrontierCode 2.0 at its default high setting. VentureBeat's headline framed the competitive picture bluntly: Opus 5.5-level coding performance at 25 percent lower API token prices — a direct shot at Anthropic's Claude coding models, which launched their own market milestone this week with Anthropic's record IPO.
The most striking number is on SWE-sweep, Google's software engineering sweep benchmark: Neon manages just 12.8 percent on the standard mini-SWE harness but more than triples to a claimed state-of-the-art 39.0 percent when running inside Google Antigravity, the company's agentic IDE — a figure Google uses to argue that the harness around a model matters as much as the model itself. Neon is also engineered to act as a fast orchestrator directing parallel Hydrogen subagents inside Antigravity, cutting redundant file reads and handling context compaction across long sessions.
Google demonstrated the model's agentic range with showcase builds: a photorealistic 3D lab bench generated through Blender MCP, a first-person WebGL2 ski simulator written from scratch on real Mont Blanc elevation terrain with over 10,000 lines of custom shaders, and an agent swarm that turned a single image into a 100,000-cell Google Sheet running edge detection through more than 500,000 live formulas.
Pricing that undercuts the frontier
Neon costs $3 per million input tokens and $15 per million output tokens, with a 1 million-token context window. Cached input tokens run $0.15 per million — a 95 percent discount designed for agentic workflows that re-read the same context repeatedly. At its default high thinking level, Google says Neon delivers a 32 percent lower cost per task than Carbon, and at xhigh it costs 36 percent or more less per task than top competing frontier models at their highest settings.
The model debuts a refined four-level thinking ladder — spanning from a fast low-cost tier up to xhigh for the hardest engineering and analytical workloads — letting teams dial intelligence and token spend per task. Google also reports Neon ranks #2 on steerability on the new Agent Arena Index for adapting cleanly to mid-task corrections, and records its lowest fabrication rate yet in internal Antigravity testing.
Security gets first-class treatment
A notable emphasis of the launch is defensive security. On AutoBaxBench, which tests whether models generate secure code from the start, Neon outperforms Carbon across easy, medium, and hard challenges and achieves state-of-the-art on the hard tier at 76 percent. On WebExploitBench v1.2, a penetration-testing benchmark, Google claims state-of-the-art at 51.26 percent — at half the cost of competing large models.
Google is also productizing that capability: a "Gemini 4 Neon for cyber" configuration covering vulnerability finding, validation, and patching rolls out to trusted defenders in the coming days through the company's Fairwind Program.
Availability
Gemini 4 Neon is rolling out today in the Gemini API and Google AI Studio, in Gemini for business via API and the model dropdown, and to paid Google AI Pro and Ultra subscribers in both Google Antigravity and the Gemini app. One caveat surfaced by developers on Hacker News: Google Cloud's documentation indicates the model ships with abuse-monitoring prompt logging on the enterprise agent platform and offers no opt-out — a consideration for teams handling sensitive code.
With Neon, Google is making its strategy explicit: rather than selling one enormous model, it ships a family of element-named models — Carbon, Hydrogen, Argon, and now Neon — and lets the benchmarks argue about which one sits on the cost-capability frontier. For coding-heavy teams watching Anthropic's trillion-dollar debut this week, the price cut is the pitch: near-frontier intelligence at a fraction of the per-task cost.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →