ModelBrief · Gemini 4 Carbon · Topic Page
Last updated Nov 13, 2026

A Four-Model Build-Off: Carbon Finished Last in Two of Three Tests — at 1/30th of Fable's Bill

Published Nov 13, 2026 · Last updated Nov 13, 2026

AI for Mortals (Pat Simmons) put Gemini 4 Carbon up against Claude Fable 5.2, Claude Opus 5.5 and GPT-6 Astra on the same three briefs: a pixel-perfect website clone, a product launch film, and a 1940s dogfight flight simulator. Every model got the exact same prompt, ran in its own lab's tool on high effort, and each build was run once; Opus and Astra ran on November 6, Fable and Carbon on November 10–11. The author's caveat up front: these are big builds, not a proxy for everyday work.

Test 1: a pixel-perfect clone of the Agrumea site

Rebuild agrumeafarm.it pixel for pixel — physics loader, 3D jar carousel and page transitions, every asset made from scratch, run through a published clone methodology with a style-assertion gate. The ranking: Fable 5.2 first ($594.80, 533 minutes), Opus 5.5 second ($60.14, 141 minutes), GPT-6 Astra third ($90.63, 68 minutes), Carbon fourth ($10.86, 99 minutes).

Test 2: a launch film for Motion

A 30–45 second product launch film at 1920×1080, 30 fps, that has to be clearly better than Motion's own launch film — interface, pacing and story all improved. The ranking: Opus 5.5 first ($14.23, 48 minutes), Fable 5.2 second ($97.62, 248 minutes), GPT-6 Astra third ($14.55, 34 minutes), Carbon fourth ($6.82, 68 minutes).

Test 3: a 1940s dogfight flight simulator

A browser-playable air-combat simulator: a realistic 20×20 km coastline built in Blender, two historically accurate 1940s propeller fighters, a real flight model (stalls, energy management, G-force blackout), ballistics, per-part damage and enemy AI that actually dogfights. The ranking: Opus 5.5 first ($106.92, 376 minutes), Fable 5.2 second ($190.88, 243 minutes), Carbon third ($10.00, 102 minutes), GPT-6 Astra fourth ($21.25, 79 minutes). The only test where Carbon did not finish last.

The bill: all three builds, Carbon at 1/30th of Fable

Total cost across the three builds, priced at each model's published API rates: Carbon $27.68 (269 minutes), Astra $126.43 (181 minutes), Opus 5.5 $181.29 (564 minutes), Fable 5.2 $883.30 (1,024 minutes). Per test, Carbon ran 10–50 times cheaper than Fable. The author's reservations: part of Fable's flight-sim log was lost when a disk filled up, so its tokens and cost may be undercounted; and once Carbon's introductory price ends on December 1, it will cost the same per token as Opus 5.5, so much of the savings will shrink.

Our take: Carbon is a clear step up from Argon, and it is good to see Google back in the running. But for building websites, applications or 3D work, the author still points to Claude or ChatGPT. The other way round, inside the Google ecosystem — computer use, automation and knowledge work — Carbon is perfectly capable, especially before the November 30 price rise.

Reference source: this article is an original rewrite produced by this site; facts and figures were checked against AI for Mortals — Gemini 4 Carbon: No-Hype Full Review & Testing (Nov 12–13, 2026)

← Back to the Gemini 4 Carbon topic page