AI Model Reviews

Gemini 4 Neon vs Opus 5.5 and GPT-6.2 Sol: World of AI Tests

Gemini 4 Neon did not make the prettiest medieval battle in World of AI's review. The host preferred Claude Fable 5.2's detail, but reported that Neon finished sooner and at roughly half the cost on that run. The review also shows the opposite kind of result: Neon made the host's favorite Xbox controller SVG, beating GPT-6.2 Sol and Gemini 4 Hydrogen in that comparison.

That's the case for trying Neon: it can give you a strong first build quickly enough to leave time for a second. It doesn't follow that Neon is Google's most capable model for every job, or that it always costs less than Sol.

Watch World of AI's Review

Source Note: World of AI's 18 November 2026 review and transcript supply the hands-on observations. The benchmark and release sources were checked on 19 November. The episode's 03:28 Mammouth segment is sponsored. The review shows selected builds, not a repeated trial with a complete record of prompts, tokens and settings.

Where Neon Really Ranks

Neon leads some evaluations, but they measure different things. Artificial Analysis's general index is not its coding-agent index, and neither is the same as Agent Arena or Zapier's business workflow test.

EvaluationResultWhat it tells you
Artificial Analysis Intelligence IndexNeon xhigh: 61, third at launch. Ahead of Opus 5.5 (59) and Sol (57), behind Fable 5.2 and Astra.A broad benchmark win over those Opus and Sol configurations, not a result for every task. Comparator scores.
Artificial Analysis Coding Agent IndexRounded: Neon xhigh 59; Opus max 58; Fable xhigh 58.Neon ran in Antigravity CLI; both Claude models ran in Claude Code. A test of model and tooling together.
Agent Arena OverallGoogle cited second at debut. The checked board places Neon high third, behind Fable and Opus at max.A moving leaderboard. The launch rank shouldn't be presented as the current rank.
Zapier AutomationBenchNeon xhigh: 56.32% at $1.69 per task. Neon high: 54.34% at $0.88.Less than two points separate the settings on this suite, at nearly twice the cost. It tests simulated business workflows, not generic screen control.

In the coding-agent runs, Neon averaged $6.86 per task versus $12.61 for Opus and $15.47 for Fable. On the separate general intelligence suite, Neon at xhigh averaged about 21 minutes per task versus about 36 for Carbon at high. Those measurements describe different suites and configurations, not universal savings or a promise that your build will finish in 21 minutes.

What the Creative Builds Showed

TestWhat World of AI showed
Three.js medieval battleThe host preferred Fable 5.2's visual detail. Neon was reportedly faster and roughly half the cost in this example; the review doesn't establish an average.
Frontend designScroll-triggered motion, typography and a stronger sense of visual style impressed the host. There's no accessibility, responsiveness or conversion comparison.
One-shot shooter prototypeThe Antigravity build had a minimap, movement controls, textured buildings and reload animation. Some areas still looked plain; the host thought Opus could add polish, but did not show a matched Opus run here.
SVG controller and motionThe host picked Neon's Xbox controller over Sol and Hydrogen for shape and gradients, and preferred Neon's detail in a separate Opus comparison. The animated cycling pelican shows a different skill: code-driven motion.
Tokyo in the rainAn Antigravity run at xhigh reportedly took about 39 minutes. Reflections and atmosphere stood out, but even a fast model can take a while on a substantial scene.
Minecraft-style sandboxIn the Gemini app, the host got shaders and water behavior after two tries and an error fix. He couldn't find the requested survival mode. The build looked convincing before it was feature complete.

Cost, Thinking Levels and Access

Google's standard paid API rates are $3 input, $0.15 cached input and $15 output per million tokens. Output billing includes thinking tokens. By comparison, OpenAI lists Sol's standard short-context rates at $2 input and $10 output. Carbon is also $2/$10 through 30 November, before rising to $4/$20 on 1 December. A lower benchmark cost for a specific Neon setup isn't the same as a lower unit price.

Neon has a shared 1,048,576-token window and accepts text, PDFs, images, audio and video; the API returns text. The animations in the review are code-built artifacts, not native video outputs. The four thinking settings are low, medium, high (default) and xhigh. Start with high for an involved build; try medium when you need quick feedback, and raise effort when there's a specific problem the lower level didn't solve.

Use gemini-4-neon in the paid Gemini API or try Google AI Studio. Google's rollout lists AI Pro and Ultra for the Gemini app and Antigravity, with separate business access. The video discusses an alleged earlier description tying Neon to Hydrogen or a Flash tier. Google's published model card says Neon is adapted from Carbon; the earlier wording isn't proof of the shipped model's lineage.

What Antigravity Adds

Google Antigravity supplies the workspace and tools to read a project, edit files and run checks. /swarm can split independent tasks across agents. Orchestrator can instead pair Neon for planning with Hydrogen for routine execution. A multi-agent run is a different experiment from a single Neon run. For a new project, establish the single-agent baseline first, then try parallel work on clearly separate tasks with one review before integration.

Computer use needs its own test. Google lists computer use as supported, but Zapier's workflow scores don't establish that Neon beats Sol at clicking through desktop software. To compare them, start with a read-only task on a test account, such as finding a report and returning its date and source. Include a changed screen or a missing row, and see whether the model recovers or asks for help. Keep live writes behind human approval.

Creator Demos and Google's Ski Simulator

Vamsi Batchu shares GeokenAI's code-built adaptation of John Keats's To Autumn, shown at 14:30 in the review. The other clip shows how a cost comparison can flip: thehype. reported spending $12.93 on Neon and $11.19 on Fable for a voxel scene, while preferring Fable's composition. That's a different prompt and run from the medieval battle, not evidence of a fixed cost gap in either direction.

Vamsi Batchu shares the To Autumn animation

Open the original post on X

thehype. compares Neon and Fable on a voxel scene

Open the original post on X

Accounts linked in the video: @intheworldofai, @IndependentEco, @chetaslua and @notjazii.

Watch Google's Chamonix ski simulator demo

This is Google's Antigravity launch showcase, not an independent timed World of AI run. The example uses Alpine elevation data and WebGL2. Pay attention to how the controls and terrain behave together, not just the landscape.

Watch on YouTube or see the full set of Google launch demos.

What to Test First

Try Neon on a job where you expect several rounds of feedback: a landing-page section, an animation or a multi-file feature. Run the same brief on your current model with the same assets, then record time and cost to an accepted result, including fixes. For a visual build, check mobile layout, keyboard use and whether the key interaction actually works. A cheaper initial run costs more if it takes several retries to reach the same standard.

Video Chapters

TimeTopic
00:00Introduction
00:38Launch benchmarks
01:00Agent Arena ranking
01:12Neon and Antigravity
01:30Context and technical specs
01:46Pricing
02:12Where to use Neon
02:25/swarm and Antigravity
03:07Thinking levels
03:28Mammouth sponsor segment
04:30Artificial Analysis results
05:58Reported changes to the model description
06:38Speed impressions
07:08Neon vs Fable 5.2: medieval battle
08:09Frontend design
09:12Call of Duty-style game
10:38Computer use
12:04SVG tests
13:11Neon vs Opus 5.5: SVG
13:25Motion graphics
14:05Tokyo rain scene
14:30To Autumn short film
17:43Ski demo
18:09Minecraft-style sandbox
19:09World of AI verdict
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