AI Model Reviews

Gemini 4 Carbon: Three Real Builds, One No-Hype Verdict

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.

The useful answer

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

TestPat's rankingCarbonFableOpusAstra
Agrumea website clone1. 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 film1. 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 sim1. 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 totalCarbon 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.

BuildGemini 4 CarbonFable 5.2Opus 5.5GPT-6 Astra
Agrumea cloneOpen buildOpen buildOpen buildOpen build
Motion launch filmOpen buildOpen buildOpen buildOpen build
1940s dogfightOpen buildOpen buildOpen buildOpen build

Video Chapters

TimeSegment
00:00The comparison
00:44Agrumea clone
15:49Website rankings and costs
17:50Carbon benchmarks and pricing
22:25Google demos
25:32How to access Carbon
25:56Motion launch film
37:12Film rankings and costs
38:491940s flight simulator
48:51Sim rankings and costs
50:06Final verdict
52:17Bonus 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.

Common questions

Did Gemini 4 Carbon win any of Pat Simmons's three builds?
No. Pat ranked Carbon fourth on the Agrumea website and Motion film, and third on the dogfight simulator. It was the cheapest API-equivalent run in each test.
Were the cost figures actual API bills?
No. Pat ran the models on subscriptions and priced logged token usage at public API rates for comparison. The Fable flight-sim log is incomplete, so its quoted cost may be an undercount.
Is the $2 / $10 Gemini 4 Carbon API price permanent?
No. Google's pricing page lists $2 per million input and $10 per million output tokens through November 30, 2026, then $4 and $20 starting December 1. Check current pricing before budgeting a run.
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