
Tech • AI • Robotics • Game
⏱ ≈ 3 min read · 21-min video — 18 min saved
Google’s Gemini 4 Neon has entered the top tier of AI models by pairing near-frontier benchmark performance with lower prices, faster output, and stronger agent workflows.
Gemini 4 Neon ranks third overall on Artificial Analysis Intelligence Index and second on Vellum’s index at 70.9%. It also leads or places near the top on legal research, tax-agent tasks, finance, accounting, spreadsheets, Automation Bench, BrowseBench, and Agent Arena, indicating broad strength across reasoning and applied enterprise work.
Pricing is listed at $3 per million input tokens and $15 per million output tokens, with a 95% discount on cached inputs. That puts Neon below leading high-end rivals on cost while remaining competitive on quality, making it notable for buyers weighing performance against inference spend.
Compared with Gemini 4 Carbon, Neon reportedly cuts cost per task by 17% and finishes tasks about 40% faster, reducing completion time from 36 minutes to 21 minutes. Output speed nearly doubled from 57 to about 110 tokens per second, a major gain for coding, document work and interactive agent tasks.
On the Artificial Analysis coding agent index, Neon scored 59, roughly alongside Claude Fable 5.2 and Opus 5.5, but at lower task cost. Reported per-task costs were $6.86 for Neon, versus $12.61 for Opus and $15.41 for Fable. In visual reasoning and presentation quality, Neon reached 1,700 Elo, topping the previous leader by 45 points.
The model includes a 1 million-token context window and four thinking levels, letting users trade off speed, reasoning depth and cost. It also shows improved instruction following and lower hallucination rates, addressing a persistent criticism of earlier Gemini releases.
Google is steering users toward Anti-Gravity, an agent harness designed to unlock Neon’s multi-step workflow abilities. Its swarm feature lets multiple agents work in parallel, splitting complex jobs into smaller coordinated tasks instead of relying on one sequential agent. That architecture appears central to Neon’s strong showing on Agent Arena and automation benchmarks.
Neon is emerging as a serious contender for computer-use and automation tasks, where models must read screens, choose actions and recover from errors. Its mix of visual reasoning, speed and instruction following may help it challenge models such as GPT 6.2 Soul, although published benchmarks do not yet fully measure graphical desktop control in real-world conditions.
Early tests show Neon improving substantially in front-end design, typography, scroll effects and overall visual taste. It also appears especially strong in SVG generation, outperforming comparisons against GPT 6.2 Soul, Gemini 4 Hydrogen and Opus 5.5 on structured prompts such as rendering an Xbox controller with accurate gradients and geometry.
Neon has shown the ability to generate detailed Three.js scenes, browser games and animated visuals from a single prompt. Examples included a Call of Duty-style prototype with lighting, textures, minimap and weapon handling, a Tokyo rain scene built in about 39 minutes, and a fully animated short inspired by John Keats’ poem To Autumn using only vanilla JavaScript.
Observers noticed that a pre-release description reportedly said Neon was built on Gemini 4 Hydrogen through reinforcement learning and offered flash-tier speed and cost, language that was later removed. That has prompted speculation about whether Neon is effectively a flash-class model marketed differently, though no definitive explanation has been confirmed.
Gemini 4 Neon signals a significant advance for Google in frontier AI, especially where buyers care about the balance of intelligence, speed and cost. The next test is whether those benchmark gains hold up across broader real-world deployment and sustained agent use.
Ask a question