Independent Benchmarks for Gemini 4 Neon: 110 Tokens/s, 65% on Cursor Bench, and Real Codebase TestsNote · digest for November 21
High-speed graphic render of code compilation and compute optimization
Note3 sections≈ 4 min

Independent Benchmarks for Gemini 4 Neon: 110 Tokens/s, 65% on Cursor Bench, and Real Codebase Tests

You drop a screenshot of a messy model list into chat, add one lazy sentence, and five minutes later you get a clean refactor with tests. No giant spec, no half-hour wait. That's how Gemini 4 Neon performed in early live tests after its November 18 release—as long as you don't overpay for the heaviest mode.

Speed used to cost you intelligence. As Julian Goldie put it: "Fast AI or smart AI? For years you had to pick one: smart almost always meant slow." In the Gemini 4 lineup, Neon sits between light Hydrogen and heavy Carbon—and independent benchmarks show you may no longer have to choose.

Where Neon beats Opus 5.5 and Carbon—and where it trails

Benchmark tracker Artificial Analysis ran Neon through its suite against top rivals.

MetricGemini 4 NeonGemini 4 CarbonClaude Opus 5.5
Intelligence Index61 (Extra High)60 (High)under 60
Coding Agent Index59 (58.6 Google)—58
Cursor Bench65.4%—58%
Speed / task time~110 tok/s (~21 min)~57 tok/s (~36 min)—

Overall, Neon took third place behind GPT-6.2 Astra and Claude Fable 5.2 at max settings, beating Carbon, Opus 5.5, and GPT-6.2 Sol. Its lead on Cursor Bench stands out most: that test measures real multi-file coding sessions inside the Cursor IDE rather than synthetic puzzles. Neon also topped Opus 5.5 on Automation Bench (+13 pts), Terminal Bench Science (+7 pts), and Agent's Last Exam (+4 pts).

It's not a knockout everywhere, though. Anthropic's model still leads by half a point on Terminal Bench and one point on Frontier Code.

Why High beats Extra High on value

Neon offers four thinking levels that control how many internal reasoning steps it takes before answering. Turning the dial to max is tempting, but the Artificial Analysis chart cools that urge quickly.

Artificial Analysis benchmarks for Gemini 4 Neon: scores and task cost across four modes vs. Carbon, Opus 5.5, and Fable 5.2 (Source: artificialanalysis.ai)

LevelIndexCost per taskWhen to use
Low44$0.75Minor edits and quick questions
Medium55$1.90Daily routine and simple features
High (default)60$3.74Main working mode for coding
Extra High61$6.44Only the gnarliest bugs

Stepping from High to Extra High buys just one extra point while raising the bill by nearly three-quarters. Default High captures almost the model's entire ceiling at half the cost of Carbon and nearly a third the cost of Opus 5.5 (while Fable 5.2 climbs past $20). The Google Antigravity blog adds that tasks on High cost roughly a third less than Carbon while running nearly twice as fast.

Live coding on Scarlet, polite OBS control, and model limits

Matthew Berman tested Neon on his open-source AI video editor, Scarlet (MIT). By then, ThePrimeagen on X was calling Neon's lead insane, while Theo wondered why such a strong model launched so quietly.

Its biggest strength is handling underspecified prompts—short, lazy requests with no detailed spec. From one screenshot and a single line, Neon compacted a model list into one row, added monospace IDs, pushed disabled items to the bottom, and wrote tests in five minutes. Asked for a mockup, it built an interactive preview right in chat in 90 seconds; from two screenshots, it added toolbar tooltips and drag-and-drop. In 45 minutes, it also wired up a Gemini provider through the Antigravity CLI over the open Agent Client Protocol (ACP), which connects terminal agents to editors. Berman summed it up: "I gave it the laziest prompts possible, and it still knew exactly what I meant." Unlike Astra, Opus 5.5, and Fable, Neon doesn't need to be told to keep summaries short.

In Computer Use mode with OBS, Neon took 2.5 minutes to spot why video recorded at 29.97 fps instead of 60 fps (the main profile was active instead of screen) and fixed the settings. When asked to revert the frame rate mid-recording, it stopped itself: "Both OBS windows are already recording, so I didn't touch or restart OBS to avoid ruining the take."

Its limits are clear too. On flashy 3D demos and games, Neon sometimes trails rivals (though it built a 3D Rubik's cube in 20 minutes). Fast output doesn't eliminate long loops on huge tasks, which is why Google keeps Carbon around for marathon projects. Finally, Antigravity's claim of near-zero hallucinations rests on internal Google tests without outside audits—Goldie recommends testing it on your own workloads. With a 1M-token context window, Neon is already live in the Gemini API, AI Studio, Antigravity, GitHub Copilot (Pro+, Business, Enterprise), and the Gemini app (AI Pro and AI Ultra), replacing Argon as Gemini 3.1 Pro retires.

Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 5