Gemini 4 Neon made the smoother game. Claude Opus 5.5 made the game Peter Yang wanted to keep looking at. That split explains more than a single model ranking: the build with more atmosphere, richer destruction and more to do also lagged during play. Which result is better depends on what you're shipping.
Across Yang's five tests, Opus 5.5 was his pick for 3D video, games and visual design. Neon narrowly won his essay comparison, and both held up under pushback, though he preferred Opus's original advice. For writing and everyday collaboration, this was the first Gemini model he felt could stand in for Opus.
Watch Peter Yang's Comparison
Source: Peter Yang's November 19, 2026 video and transcript. These are one creator's judgments on the outputs he shows, with Gemini 4 Carbon as a third reference. The video doesn't establish how often the same results would recur. Product details were checked against Google and Anthropic on November 20.
Five Tests at a Glance
| Task | Yang's pick | What made the difference |
|---|---|---|
| Shibuya travel ad | Opus 5.5 | A wider opening shot, a fuller crowd and a more convincing move down to street level. |
| 3D monster game | Opus for spectacle, Neon for smoothness | Opus offered more varied attacks and destruction. Neon ran more smoothly on Yang's machine. |
| Website redesign | Opus 5.5 | A bolder visual concept and stronger motion. Neon had a clear call to action, but more copy than Yang wanted. |
| Personal essay | Neon, narrowly | Yang liked the opening and the pacing. All three essays were close. |
| Advice under pushback | Close; slight Opus preference | All three mostly stood by their advice. Opus left room for the interviews Yang enjoys making. |
Where Opus Still Wins
The Shibuya film. Neon got the wet-street reflections right, but its junction looked small and some pedestrians slid rather than walked. Opus opened farther back, with a train and a fuller crowd, then descended to street level. The people still looked artificial, but the camera move sold the scale. Yang puts both runs at around 40 minutes and Carbon at 73 minutes. Those are timings for these builds, not general latency figures.
The monster game. Opus staged the entrance like a film, with a roar and a building breaking apart. The richer effects and destruction also made it lag noticeably. Neon ran more smoothly but looked simpler, and one breath attack appeared to come from the monster's neck. In Carbon's build, the camera moved without letting Yang see the character move properly. For a playable prototype, test input response and frame pacing separately from the opening sequence.
The Behind the Craft homepage. Neon gave the offer a clear button and used an animated reel, though Yang thought the page packed in too much copy. Opus committed to a distinctive clothes-hanger-and-drawers concept with stronger motion, closer to the brief's request for a bold visual direction. Yang didn't test conversion rates. A more memorable homepage isn't necessarily a better-performing one.
Where Neon Catches Up
Writing: Each model wrote a fictional reflection from a software engineer who no longer writes code by hand. Yang chose essay A (Neon), then C (Carbon), then B (Opus). Neon opened with a familiar keyboard shortcut used in the wrong context, a concrete detail that did more for him than a neatly packaged ending. He later disclosed he already recognized Carbon's essay, though he hadn't identified Neon or Opus. It's a narrow preference on one prompt, not a fully blind three-way win. In a separate revision turn, he found Neon's before-and-after suggestions easier to scan than Carbon's longer response.
Advice: Neon proposed moving interviews to a second channel. Opus suggested more timely solo tests while keeping interviews tied to relevant launches. Carbon wanted to drop talking-head interviews entirely. Opus fit Yang's preferences a little better, but all three mostly held their ground when challenged. Yang reports the clearest concessions followed real data. The useful test is whether a model can reconsider when the evidence changes, without reversing itself just because you disagree.
Benchmarks, Pricing and Access
Google DeepMind's evaluation lists Neon ahead on CursorBench 4.1 (65.4% vs 58.1%) and slightly behind on Terminal-Bench 4.1 (65.2% vs 65.8%). The report draws on different sources and setups for some results. These scores measure particular coding tasks; they don't tell you which game will run better or which homepage you'll prefer.
Google positions Neon between Hydrogen for fast everyday execution and Carbon for harder, long-running work. Yang would try Neon first for most Gemini work, use Hydrogen for quick questions, and reserve Carbon for jobs such as a large audit or migration. He didn't test those jobs here. Carbon's weaker showing on these visual builds is a reason to test it on your own workload, not to write it off.
At the listed standard API rates, Neon is $3 per million input tokens and $15 per million output tokens; Anthropic lists Opus 5.5 at $4 and $20, respectively. That's 25% lower per standard input or output token, not necessarily 25% less per finished task. Token use, retries and editing time can change the result. These are API prices, separate from subscription limits. Google lists Neon on AI Pro as well as Ultra; its launch also includes the paid Gemini API and Antigravity for Pro and Ultra subscribers. Check your country and plan before subscribing.
Keep Your Skills Portable
Yang keeps his reusable skills in a private GitHub repo so his advisor's instructions can move between models. Try the same approach with one workflow: write the goal, inputs, output format, decision rules and a few examples in plain Markdown. Keep tool-specific instructions separate. Leave credentials and private source material out of the skill file.
Then run one real task in both tools using the same brief and rubric. For writing, hide the model names and count how much editing each draft needs. For a game or website, try the controls and test the awkward paths, not just the first screen. That gives you a reason to switch models without having to rebuild the workflow each time.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | Why Gemini 4 Neon is getting so much hype |
| 00:42 | Carbon vs Neon vs Hydrogen: Which Gemini 4 model to use |
| 02:28 | Test 1: Shibuya Crossing 3D video ad |
| 03:57 | Test 2: 3D monster game |
| 06:05 | Test 3: Website redesign |
| 07:50 | Test 4: Blind writing test |
| 12:39 | Test 5: Holding its ground on advice |
| 14:30 | Verdict and portable AI skills |