Google Unleashes Gemini 4 Neon, a Cheaper Frontier Brain for Daily Work
Google rolled out Gemini 4 Neon on Nov. 18, 2026 — a faster, cheaper cut of its flagship Gemini 4 Carbon built for everyday production work. Koray Kavukcuoglu announced the model, which carries a 1M-token context window and tops several coding, security and long-context benchmarks. The catch: it's paid-tier only.
- Original source: note.com.
- The source published it on Nov 18, 2026 at 7:59 PM.
- No second outlet has reported it yet. Treat it as a single-source claim.
Google's model treadmill keeps spinning. On Nov. 18, 2026, Koray Kavukcuoglu announced Gemini 4 Neon, a frontier-class model tuned for what the company calls high-frequency daily workflows — complex software engineering, knowledge work and scientific tasks run end-to-end. It's derived from the flagship Gemini 4 Carbon, keeping Carbon's reasoning quality while running faster and cheaper.
The spec sheet is broad: a 1M-token context window covering input and output, up to 64K synchronous output or 1M in background mode, multimodal input across text, image, audio and video, a March 2026 knowledge cutoff, and four thought levels from low to xhigh. Model ID is gemini-4-neon, listed on OpenRouter as google/gemini-4-neon.
On Google's own model card, Neon scores 61 on the AA Intelligence Index v5.0.1 at xhigh — third overall out of 118 models, ahead of Claude Opus 5.5 at 59 and GPT-6.2 Sol at 57. It hits 70.9% on Vals Index v2.2, takes first on AutomationBench at 56.3% against Opus's 42.5% and Sol's 41.3%, and tops Vending-Bench 2 at $16,843. Google notes all figures are official and pre-verification.
Coding is where it flexes hardest. It leads the AA Coding Agent Index at 58.6 and CursorBench 4.1 at 65.4%, while SWE-sweep jumps to 39.0% on the Antigravity harness — more than triple Opus's 10.1% and Sol's 8.3% — and still leads at 12.8% on the mini harness. It doesn't sweep everything: Terminal-Bench 4.1 lands at 65.2% behind Opus's 65.8%, and FrontierCode 2.0 comes in at 46.2% against Opus's 47.3%.
Elsewhere it's a wall of firsts: RiemannBench at 84.0%, Terminal-Bench-Science 0.1 at 70.5%, HLE-Diamond at 71.3%, GraphWalks at 95.2% and LVBench at 89.0%, plus the Agent Arena Index at 1336 across 4,200+ sessions. On security it posts 90.5% on SEC-bench Pro at roughly half the price of rivals, 70.7% on CWE-Bench v2 — up 3.4 points on Carbon — 51.3% on WebExploitBench v1.2 and 76% on AutoBaxBench's hard tier. Surge AI evaluations put it first in most finance, accounting, spreadsheet modeling and complex constraint compliance categories.
Pricing sits between the competition: $3.00 per 1M input tokens, $0.15 per 1M cached input — a 95% discount — and $15.00 per 1M output tokens, with long decoding supported up to 1M. OpenRouter's batch version runs $1.50 and $7.50, while Vertex AI regional pricing runs $3.30 and $16.50. The drawback: Neon is only available through paid billing surfaces like Antigravity or Google AI Studio, and its parent Carbon remains the base for the hardest long-horizon projects. Argon, meanwhile, handles coding and cyber defense with a 12-month GA, and Hydrogen covers high-speed daily execution with AA Index scores of 51 at high and 54 at xhigh. Antigravity's Orchestrator flips between Neon as planner and Hydrogen as executor inside a single session, repairing the plan when Hydrogen stalls. Google's blog showed off demos including a 10,200-line WebGL2 ski simulator using real Mont Blanc and Chamonix elevation data, a 3D recreation of Turing's 1936 study, and a 500,000-formula, 100,000-cell Google Spreadsheet generated from a single image.
TopicsGoogle · Gemini 4 Neon · Koray Kavukcuoglu · Gemini 4 Carbon · Claude Opus 5.5 · GPT-6.2 Sol · Google Antigravity · OpenRouter
Source: note.com
Written by AI from verified reporting.
More stories
- Tech & AI x.com 1 viewRushed AI-spurred software patches likely behind recent tech outages
- Tech & AI coinpedia.org 1 viewKress Tells Nvidia Staff 'Let's Move Forward' at First All-Hands Since Huang Left
- Tech & AI cryptobriefing.com 1 viewChina's military hunts Nvidia chips as Washington tightens export rules
- Tech & AI hindustantimes.com Caught by @pz129313 1 viewKalanick Prepares to Lead Protest to Reinstating Nvidia CEO Jensen Huang

