Google launched Gemini 4 Carbon on 10 November 2026. Rather than infer how it builds from a leaderboard, Pat Simmons gave it the same three briefs as Claude Fable 5.2, Claude Opus 5.5 and GPT-6 Astra: recreate a highly animated commerce site, produce a launch film, and build a browser flight simulator. He published the complete prompts, token logs, rankings and live builds.
In Pat's three first-run builds, Carbon did not take first place. It placed fourth, fourth and third, but was the least expensive in API-equivalent tokens every time. That makes it interesting for cost-sensitive prototypes and Google-native workflows, not an automatic winner for visual fidelity. One run per model and different agent harnesses make these illustrative comparisons, not a controlled model benchmark.
Watch the Full Review
What Pat Actually Tested
Each model received the same prompt for each task and access to broadly similar supporting tools, including Blender, an image generator and sound effects. Carbon ran in Google Antigravity; Astra in Codex; Fable and Opus in Claude Code. Pat set each to high effort and judged the finished artifact, not merely whether the code compiled. His clone-app playbook and launch-video skill are linked in the original research page.
Those conditions are useful, but they are not identical lab environments. Each harness can change tool use, iteration and context management. Opus and Astra ran on 6 November; Fable and Carbon on 10-11 November. There was one run per model per brief. Pat's ranks are his subjective assessment of those outputs, not a reproducible population estimate. For a fair buying decision, inspect the live builds and run your own task in the same environment.
Three Tests, Four Models
| Test | Pat's ranking | Carbon | Fable | Opus | Astra |
|---|---|---|---|---|---|
| Agrumea website clone | 1. Fable 5.2 2. Opus 5.5 3. GPT-6 Astra 4. Gemini 4 Carbon | $10.86 | $594.80 | $60.14 | $90.63 |
| Motion launch film | 1. Opus 5.5 2. Fable 5.2 3. GPT-6 Astra 4. Gemini 4 Carbon | $6.82 | $97.62 | $14.23 | $14.55 |
| 1940s flight sim | 1. Opus 5.5 2. Fable 5.2 3. Gemini 4 Carbon 4. GPT-6 Astra | $10.00 | $190.88* | $106.92 | $21.25 |
| Three-build total | Carbon cheapest; no overall quality winner assigned | $27.68 | $883.30* | $181.29 | $126.43 |
USD, API-equivalent estimates from Pat's token logs, although the builds ran on subscriptions. *Part of Fable's flight-sim log was lost when Pat's disk filled, so its quoted cost and total may be too low. Read the full per-build token and time receipts.
Test 1: The Agrumea Website Clone
The brief was to rebuild Agrumea Farm, including the physics loader, 3D product jars, carousel, page transitions and interactions. Pat ranked Fable first for fidelity, followed by Opus, Astra and Carbon. Carbon's 99-minute, $10.86 run is a useful price signal, but its fourth-place result says the cheapest first draft was not the best clone. Fable's first-place version took 533 minutes and $594.80 in API-equivalent usage, a striking quality-cost trade-off. Examine the Carbon build next to the Fable build, including product-page behavior, not just the hero screenshot.
Test 2: A Launch Film for Motion
For a 30-45 second product film inspired by Motion, the models were asked to study the reference, make graphics and render a finished video. Pat placed Opus first, then Fable, Astra and Carbon. Carbon's film cost $6.82 API-equivalent and took 68 minutes, but was Pat's least preferred of the four. Opus's leading result was $14.23 and 48 minutes. This is the clearest case where another model's better judged result was available at a relatively small extra token cost. Review the Carbon film and Opus film with sound; typography, pacing and the clarity of the product story matter more than a still frame.
Test 3: A 1940s Dogfight Flight Simulator
The hardest software brief called for a browser-playable propeller-plane sim over a coastal landscape, with controls, enemies, a full sortie and a verification pass. Pat ranked Opus first, Fable second, Carbon third and Astra fourth. Carbon beat Astra on this test while costing less in Pat's estimate ($10.00 versus $21.25). Its run was stopped and resumed in the same conversation, so the 102-minute figure reflects that workflow. Try the Carbon simulator and Opus simulator yourself. A convincing menu or cinematic flyover is not proof the flight model or enemy AI works through a complete mission.
Pat also let Carbon work on an Eiffel Tower reconstruction prompted by Google's own demo. At the 52:17 bonus, he shows an intermediate result after roughly six to seven hours. It is a separate exploratory run, not a fourth scored comparison.
Carbon Pricing and Access: Check the Date
Google's launch announcement says Carbon is available through the API to paying developers and is rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode. Google reports strong coding and agent benchmarks, but those vendor evaluations ask different questions from Pat's three visual production tasks. Google also presents its Eiffel Tower and other multi-agent demos as examples of what Carbon can do; they do not imply a one-pass reproduction will match the showcase.
Google's published API pricing lists $2 per million input tokens and $10 per million output tokens through 30 November 2026, rising to $4 and $20 from 1 December 2026. Carbon has no free API tier in that pricing table, and thinking tokens count as output. Other tools, rendering, image generation and audio may cost extra. Pat's $27.68 three-build comparison uses the introductory rate; it should not be projected unchanged into December or treated as an invoice for his subscriptions.
Open All 12 Live Builds
These are Pat's published artifacts, grouped by task. The original article holds the complete prompts, screenshots and token receipts beside them.
| Build | Gemini 4 Carbon | Fable 5.2 | Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| Agrumea clone | Open build | Open build | Open build | Open build |
| Motion launch film | Open build | Open build | Open build | Open build |
| 1940s dogfight | Open build | Open build | Open build | Open build |
Video Chapters
| Time | Segment |
|---|---|
| 00:00 | The comparison |
| 00:44 | Agrumea clone |
| 15:49 | Website rankings and costs |
| 17:50 | Carbon benchmarks and pricing |
| 22:25 | Google demos |
| 25:32 | How to access Carbon |
| 25:56 | Motion launch film |
| 37:12 | Film rankings and costs |
| 38:49 | 1940s flight simulator |
| 48:51 | Sim rankings and costs |
| 50:06 | Final verdict |
| 52:17 | Bonus Eiffel Tower build |
Sources, Rights and Scope
This article draws on Pat's video and transcript, his published comparison with complete prompts, Google's release and the official API price table, checked on 12 November 2026. We summarize rather than reproduce Pat's full prompts; use his original page for exact wording. A clone in a comparative test is a demonstration, not permission to publish another company's design, imagery or trademarks as your own product.