Introducing Gemini 4 Neon

Glass Epoch DeskOriginal digestSource: Google DeepMind

AI-assisted digest, checked against the original source before publishing. How we work

Google DeepMind launches Gemini 4 Neon, a frontier model tuned for fast end-to-end agent task completion across coding, knowledge work, and science.
Introducing Gemini 4 Neon
Photo: Levi Stute · free stock license

On the Artificial Analysis Intelligence Index, Neon scores 61 at its xhigh thinking level, ranking #3 overall and ahead of comparable models while trailing larger, slower systems. It places #2 on the Vals Index at 70.9%, #2 on steerability in the new Agent Arena Index — reflecting clean adaptation to mid-task user corrections — and records its lowest fabrication rate yet in internal Google Antigravity testing. Pricing is set at $3 per million input tokens, $15 per million output tokens, with cached inputs at $0.15 per million (a 95% discount) and a 1 million-token context window that holds spreadsheets, PDFs, charts, audio, and video together.

Benchmark results span several domains. Neon takes #1 on the Legal Research Bench and Tax Agent Bench, leads most Surge AI benchmarks across finance, accounting, and spreadsheet modeling, scores 73.5% on Chartography, and performs on par with leading models on Agents' Last Exam v1.1 at 75.8%. It also ranks #1 on Zapier's AutomationBench (56.3%) and Andon Labs' Vending-Bench 2 ($16,843), debuts #2 on Arena's Agent Arena Index with 1336 across 4,200+ real-world agent sessions, and is best in class on Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%).

For developers, Neon ranks #1 on the Artificial Analysis Coding Agent Index (58.6), with 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default high setting. On SWE-sweep it leads the mini-SWE harness at 12.8% but reaches 39.0% inside Google Antigravity — a gap DeepMind attributes to what the Antigravity harness unlocks. The model is built for long-running, tool-heavy sessions: it reduces redundant file reads, shares progress as it works, compacts context reliably, and can act as a fast orchestrator directing parallel Hydrogen subagents. DeepMind also showcased demos, including a Blender MCP photorealistic 3D lab bench and a from-scratch WebGL2 freeride ski simulator with 10,200 lines of custom shaders and 120 Hz physics.

The open questions center on whether the harness-dependent gains, particularly the SWE-sweep jump from 12.8% to 39.0%, hold for teams running Neon outside Google's own environment, and whether the steerability and fabrication claims translate into production reliability at the stated price point as enterprises push longer autonomous agent runs.

Key facts

  • • Neon scores 61 on Artificial Analysis Intelligence Index, ranking #3 overall.
  • • Priced at $3 per million input and $15 per million output tokens.
  • • 1 million-token context window spanning documents, charts, audio, and video.
  • • #1 on Artificial Analysis Coding Agent Index with a score of 58.6.
  • • SWE-sweep score rises from 12.8% on mini-SWE to 39.0% in Antigravity.

Why it matters

Builders gain a frontier agent-capable model priced below typical frontier rates, with cached inputs at a steep discount that makes long-context, tool-heavy sessions economically viable. The large gap between Neon's baseline and Antigravity-scaffolded benchmark scores signals that harness engineering may matter as much as model choice for production agent workloads. Expect teams to scrutinize whet

Spotted an error? Send a correction.

Get the day's AI news in one email

The Daily Epoch · 60-second read · unsubscribe anytime

Related

Models
Google DeepMind2 min read

Introducing Gemini 4 Hydrogen

Google DeepMind's Gemini 4 Hydrogen is a speed- and efficiency-focused model aimed at everyday coding and agentic workflows, alongside a new Antigravity Orchestrator that routes work between models.