Join the LuminaBench Discord
LuminaBench
← Lumina Journal

Gemini 4 Neon is now out

Google’s new model is strong at coding. Here are some of our demos with it.

Google
  • Gemini 4 Neon
Rolling out in the Gemini API and Antigravity
Original Discord card ↗

Neon came in at 65.4% on CursorBench 4.1 at Extra High, top of the board at launch. High got 63.7% and cost $6.54 per task against $10.67 at Extra High. It also scored 61 on Artificial Analysis at Extra High. 1 2 3

I tried Extra High in Antigravity. The seaplane felt quick to build, and the flight UI is much cleaner this time. 4

I like the design on the pagoda too. This one’s worth watching through the fly-around. 5

One thing in Google’s chart is easy to miss. Neon scored 12.8% on SWE-sweep in Mini-SWE and 39.0% in Antigravity. The harness makes a big difference here. 3

The November comparison from Google DeepMind’s Gemini 4 Neon model card.Source: Original model cardCredit: Google DeepMind
Google compares Gemini 4 Neon, Claude Opus 5.5 and GPT-6.2 Sol across coding, work, reasoning, science, multimodal and long-context tasks.
Open original image ↗

Sources & references

  1. Google’s Gemini 4 Neon announcement ↗

    18 November. Covers the rollout, model settings and SWE-sweep harnesses.

  2. CursorBench 4.1 ↗

    Cursor’s changelog dates 4.1 to 18 November. Cost per task is calculated from API token rates.

  3. Google DeepMind’s Gemini 4 Neon model card ↗

    The supplied comparison image matches the November results table.

  4. Our Gemini 4 Neon seaplane ↗
  5. Our Gemini 4 Neon voxel pagoda ↗