OpenAI says GPT-6.2 Astra beats Claude Fable 5.2 for a fraction of the cost. Astra and Fable both cost $8 / $40 per million tokens, so the bill depends only on how much each model writes. Astra and its cheap sibling GPT-6.2 Sol got seven prompts, Fable the two 3D ones. One prompt, one answer, no edits. Everything runs in your browser.
Kumo, a fictional sparkling green tea in slim cans: the can drawn in code, scroll animations, a flavour picker that recolours the page.
Cream and serif, a yuzu can on a yellow card, three flavour pills. Round winner.
A dark green hero with a glowing can, big flavour buttons that recolour everything, more sections and motion.
Northbeam Advisory and a mock survey of 120 small businesses, charts drawn in code, ending on a 90-day plan.
A 120-square grid, a blocker chart and one example per recommendation.
Same numbers, all correct, slide counter and keyboard navigation. Round winner on the bill (quality tie).
From Venice, two people, late June, a budget of 4,000 euros: map drawn in code, itinerary, nights, drive times, budget, bookings.
The prettier map, with the red route through every stop and a day strip.
Bases in the header, day-by-day list, a budget ring that adds up to 4,000 euros and a book-ahead checklist. Round winner.
BrightPath, a fictional 40-person software company: KPI cards, revenue and MRR, cash runway, a 12-month forecast with sliders, bear/base/bull.
Strong header, four clear cards, a forecast studio and the full monthly table. Round winner.
Plainer, but it flags the month the bear case runs out of cash.
A paper boat escaping down a rainy street to the sea: drains, a cat, checkpoints, a timer, rain made with WebAudio.
Bright and friendly, a cat on the kerb, drains that cost you four seconds. Round winner.
Darker and moodier; checkpoints give you time back.
A container terminal at night in the rain, drones that hunt you, three uplink relays to destroy.
The film look: wet orange containers, lamps in the rain. Round winner. From the November 10 run.
Bright for a night, painted containers, a red drone that shoots back.
Darker and very rainy, the longest file of the round.
Water with reflections, a waterfall, a rowing boat, a jetty, reeds, mist, fireflies, a time-of-day slider and rain.
Reeds in the foreground, the dock, a waterfall into the lake. Round winner. From the November 10 run.
Turquoise water, the waterfall, a boat by the jetty; misty at night in the rain.
One shader uses a reserved word (patch), so the cliffs do not draw: floating pines, no mountains.
Every token at list price, no cache discount (input / output per million): GPT-6.2 Astra $8 / $40, Claude Fable 5.2 $8 / $40, GPT-6.2 Sol $2 / $10. Everything ran on our ChatGPT and Claude subscriptions, so the extra spend was $0: the table shows what the same runs would cost on the API.
| Model | Builds | Output tokens | Bill at list price | Generation time |
|---|---|---|---|---|
| GPT-6.2 Sol | 7 | 168k | $1.95 | 56 min |
| GPT-6.2 Astra | 7 | 123k | $8.29 | 44 min |
| Claude Fable 5.2 | 2 | 350k | $20.48 | 59 min |
Same two 3D prompts only: Astra $1.98 in 14 min, Sol $0.68 in 19 min, Fable $20.48 in 59 min.
The exact prompts are for Ruocco Academy members. The Academy is not open yet: join the waiting list and you get one email when it opens. The demos on this page are free to use for everyone.
Join the waiting list