AI Model Reviews

Gemini 4 Neon Hands-On: Paul Lipsky Tests It Against Carbon

Gemini 4 Neon is officially out. In Paul J Lipsky's hands-on review, the new model didn't just finish faster than Carbon: he preferred its paintball visuals and its animated short. Carbon still won his space-fighter comparison. That's a better reason to try Neon than a blanket claim that it replaces Carbon.

Google launched Neon on 18 November 2026 and places it between Hydrogen for routine work and Carbon for the hardest, longest jobs. For a project where you need several rounds of feedback, the interesting question is whether Neon gets you to an acceptable result sooner.

Watch Paul J Lipsky's Hands-On Review

Lipsky used early access for his tests and also showed Neon appearing on a regular Pro account. The examples are his runs and judgments, not a controlled benchmark. He reports Neon was about 2.5 times as fast as Carbon in his testing, but the episode does not provide a task-by-task timing or cost log.

What Neon Did Better, and Where Carbon Held Up

TestWhat Lipsky observed
PortraitMixed. He liked depth and texture in one version and the more realistic facial details in the other. This wasn't a decisive win.
GladeNeon built a small foraging game. It needed a follow-up to fix bridge collision and stop the player clipping through the ground while foraging. No Carbon run was shown.
Paintball capture the flagSame prompt, Godot and Blender, both at high thinking. Neon got the visual edge for camera movement and sharper graphics; Carbon felt easier to play.
Space fighterCarbon edged it on polish, especially the water. Lipsky found Neon's ship controls responsive and its build playable.
90-second animated shortNeon won on cleaner character motion and more ambitious shots. Both builds changed styles as the characters passed through doors, but neither was flawless.

The paintball result is the one I'd carry into a real project: sharper graphics didn't make the game easier to play. Review the next build with two separate checks, one for how it looks and one for whether somebody can pick up the controls and complete a round.

Launch Specs and What the Benchmarks Measure

The Gemini API release notes list gemini-4-neon as generally available on the paid tier, with a 1 million-token context window and four thinking levels: low, medium, high (the default) and xhigh. The model card says Neon was adapted from Carbon and distinguishes up to 64K default synchronous output tokens from longer background generation. Neon can read text, images, audio and video and outputs text. A Blender animation is an artifact made with tools, not native video output.

In Artificial Analysis's launch evaluation, Neon at xhigh scored 61 on the Intelligence Index, versus 60 for Carbon at high. The evaluator reported about 21 versus 36 minutes per task on that suite. These are different effort levels, and the timing is a suite average, not a prediction for a game build or confirmation of Lipsky's 2.5x figure.

Google's coding evaluation is a good illustration of why the surrounding tools matter: Neon scored 12.8% on SWE-sweep in the mini-SWE setup and 39.0% in Antigravity. Those are results for two model-and-tool configurations, not two different Neon models.

Four Official Google Demos

These are the four videos from Google's launch article, not Lipsky's builds. They show what Neon can do with Antigravity and connected tools. They don't tell us how often the same brief succeeds or how much human review was needed.

Gemini Elements 3D: a Blender lab bench

An Antigravity swarm used Blender MCP to build an elemental still life. Notice the material and lighting choices: this shows a model directing a 3D tool, not returning a rendered frame directly from the Gemini API. Watch on YouTube.

Chamonix Ski Simulator: a WebGL2 game

Neon built a first-person skiing experience using real Alpine elevation data. The useful check is how the terrain, controls, carved tracks and audio behave together while the player moves. Watch on YouTube.

Turing Machine 1936: an interactive reconstruction

Google says the swarm worked from archival photos to reconstruct Alan Turing's study in Cambridge and put a working Universal Turing Machine inside the scene. The reconstruction and the simulation are two separate things to evaluate. Watch on YouTube.

Sheets CV Lab: an image processed in cells

Google turned one image into a 100,000-cell spreadsheet that runs image-processing effects with formulas rather than an embedded picture or Apps Script. It demonstrates what a spreadsheet can compute, not what most workbooks should look like. Watch on YouTube.

Pricing: Neon Is Below Carbon's Regular Rate, Not Its Launch Rate

Google's standard paid API rates, in USD per million tokens, checked 19 November 2026. Output pricing includes thinking tokens.

ModelInputCached inputOutput
Hydrogen$1.50$0.075$7.50
Neon$3.00$0.15$15.00
Carbon through Nov. 30$2.00$0.10$10.00
Carbon from Dec. 1, as listed$4.00$0.20$20.00

At the same token volumes, Neon is 50% more expensive than promotional Carbon and 25% cheaper than regular Carbon. Neither comparison tells you the cost of a finished run. Count tokens, retries, tool charges and review time, especially if you change thinking levels.

Where to Access Gemini 4 Neon

  • Gemini app: Google AI Pro and Ultra subscribers can look for Neon in the Gemini model picker as it rolls out.
  • Developers: Try Neon in Google AI Studio or request gemini-4-neon on the Gemini API paid tier. The standard API pricing page lists no free Neon tier.
  • Antigravity: Google's Antigravity rollout covers eligible paid Pro and Ultra users on desktop, CLI and IDE extensions. Orchestrator can use Neon to plan and Hydrogen to do implementation work. A standalone Neon run and an Orchestrator run are different setups.
  • Business: Google lists Neon for Gemini for business through the API and model dropdown. Check eligibility and access in your workspace.

Hydrogen, Neon or Carbon: What to Try First

Start with Hydrogen for fast, repeatable jobs. Try Neon at high when the job needs more reasoning but still benefits from quick iteration: a multi-file fix, a playable prototype or a brief built from several documents. Keep Carbon in the comparison when polish or a long run is worth waiting for, or when the introductory price changes the economics of your workflow.

Pick one task and hold the brief, tools and acceptance checks steady. For a prototype, test the first playable minute, not just the first frame. For an analysis, check its numbers against the source files. Then compare time and cost to the first result you'd actually use. That's the advantage Neon's launch is worth testing for.

Video Chapters

TimeSegment
00:00Introducing Gemini 4 Neon
01:05Access and availability
01:53Side-by-side portrait test
02:59Glade: forest foraging game
04:05Capture-the-flag paintball comparison
05:49Space-fighter game comparison
06:37Animated short film comparison
08:33Speed and choosing a model

Sources and Limits

Paul J Lipsky's review and supplied transcript support the hands-on observations. Google published the four demos in its 18 November launch article. The cited API docs and independent evaluation cover a different set of claims. JQ AI SYSTEMS did not rerun these tests. Evidence checked on 19 November 2026.

Common questions

When did Gemini 4 Neon launch?
Google announced Gemini 4 Neon on 18 November 2026. The paid Gemini API lists it as generally available under gemini-4-neon; the Gemini app and Antigravity rollout covers eligible Google AI Pro and Ultra subscribers.
Is Neon cheaper than Carbon?
At standard prices, Neon is $3 input and $15 output per million tokens, versus Carbon's regular $4 and $20. Carbon's introductory $2 and $10 rates run through 30 November 2026, so its per-token price is temporarily lower than Neon's.
Did Neon beat Carbon in Paul J Lipsky's tests?
Lipsky preferred Neon's visuals in the paintball test and its short animation. He gave Carbon the edge in the space-fighter game and found strengths on both sides in the portrait test. He reported about 2.5 times faster results from Neon in his own testing.
Share
X LinkedIn Reddit
Build Yours

Want a system
like this one?

Book a free 30-minute call. We map your situation, identify the highest-impact automation, and figure out if we are a fit.

Book Free 30-min Call