On November 19, 2026 at 05:51 Beijing time, Google (@Google) introduced Gemini 4 Neon, followed at 06:06 Beijing time by @antigravity announcing its rollout across the Google Antigravity desktop app, CLI, and IDE Extensions for paid Google AI Pro and Ultra subscribers as well as Gemini for business enterprise users.
Model Capabilities, 1M Context, and Orchestrator Pairing
Built for complex coding, knowledge work, and science workflows, Gemini 4 Neon offers a 1M-token context window and four thinking levels to control compute effort. Google reports that Neon ranks #3 overall on the Artificial Analysis Intelligence Index at xhigh, #1 on the Artificial Analysis Coding Agent Index, #2 on the Vals Index, and #2 on steerability on Agent Arena, while recording Google's lowest fabrication rate yet in internal Antigravity testing.
Inside Google Antigravity, developers can also use Orchestrator to pair Gemini 4 Neon as the planner with Gemini 4 Hydrogen as the executor, combining Neon's judgment on key decisions with Hydrogen's execution speed.

Image source: Conceptual overview diagram created by DigIntoAI based on official Google and Google Antigravity announcements.
Published Benchmark Comparison vs. Claude Opus 5.5 and GPT-6.2 Sol
Google published an 18-row benchmark table comparing Gemini 4 Neon against Claude Opus 5.5 and GPT-6.2 Sol (see our OpenAI GPT-6.2 Astra and GPT-6.2 Sol launch report), with methodology details hosted at deepmind.google/models/evals-methodology/gemini-4-neon:

Image source: Official benchmark chart published by @Google on X.
Capability Indices: **AA Index v5.0.1 scores **61 for Neon vs. 59 for Claude Opus 5.5 and 57 for GPT-6.2 Sol; Vals Index v2.2 reaches 70.9% vs. 70.1% and 66.1%.
Agentic Coding: **CursorBench 4.1 reaches **65.4% vs. 58.1% for Claude Opus 5.5; SWE-sweep (Provider harness) hits 39.0% vs. 10.1% and 8.3% (Mini-SWE harness: 12.8% vs. 3.6% and 11.6%); Terminal-bench 4.1 (cheating-adjusted accuracy) is 65.2% vs. 65.8% and 61.5%; FrontierCode 2.0 main score is 46.2% vs. 47.3% and 43.8%.
Knowledge Work, Science/Math, and Reasoning: **AutomationBench scores **56.3% (vs. 42.5% and 41.3%); Agents' Last Exam v1.1 reaches 75.8% (vs. 71.6% and 73.0%); RiemannBench hits 84.0% (vs. 80.0% and 80.0%); Terminal-Bench-Science 0.1 reaches 70.5% (vs. 63.3% and 68.0%); HLE-Diamond scores 71.3% (vs. 67.9% and 64.0%).
Multimodal Understanding, Long Context, and Cybersecurity: **MKA (Climb WSR) reaches **56.5% (vs. 53.0% and 52.7%); LVBench hits 89.0% (vs. 83.7% and 87.6%); Chartography scores 73.5% (vs. 66.3% and 75.1%); GraphWalks (256k to 1M, BFS F1) reaches 95.2% (vs. 66.8% and 80.1%); cybersecurity benchmarks score 70.7% on CWE-Bench v2, 90.5% on SEC-bench Pro, and 51.3% on WebExploitBench v1.2. Google has not yet published a public API model ID or token pricing schedule for Gemini 4 Neon.
