Give three AI models the same brief for a rainy Shibuya Crossing and you'll get three very different results. In Peter Yang's comparison, GPT-6.2 Astra delivered first, Claude Fable 5.2 made the stronger film, and Gemini 4 Carbon did its best work when the job shifted from 3D to writing. Yang kept Fable as his overall pick, but the five tests give better reasons to choose a model than a single ranking does.
Watch Peter Yang's Full Comparison
The three models received the same starting brief per task. Yang describes two attempts for the builds, including a first playtest and a revision for the game. The essay had three blind rounds; the advice test used four follow-up objections. These are one reviewer's judgments, not repeated measurements of the models under identical hardware, tools and reasoning budgets.
Five Tests at a Glance
| Test | Yang's verdict | What mattered |
|---|---|---|
| Shibuya travel film | Fable | More dramatic camerawork and a stronger street-level scene |
| Kaiju game | Fable, then Astra, then Carbon | Cinematic opening, readable controls and destruction |
| Homepage redesign | Fable for motion; Astra for clarity | How much animation the offer can carry |
| Blind essays | Carbon won 2 of 3 rounds; Fable won 1 | Specificity and fidelity to the brief |
| Advice under pressure | Fable and Carbon ahead of Astra | Scope of evidence and whether revisions were justified |
1. Shibuya Film: Fastest Was Not the Most Convincing
The 20-second travel brief called for a single camera move from the skyline to street level, with rain, umbrellas and homemade neon signage. Yang reported 11 minutes for Astra, almost an hour for Fable and 73 minutes for Carbon. Astra's wet reflections and neon worked, but its streets and buildings didn't look quite right. Carbon reached street level but felt cluttered to Yang. Fable won him over with more dramatic camera angles, Tokyo Tower in the distance and a stronger street-level scene. He did not think any build rendered crowds and traffic particularly well.
The prompt's 15-minute final render cap and the reported build durations are different measures. The video does not establish that any named run exceeded the final-render cap.
2. Kaiju Game: You Have to Be Able to Play It
In the kaiju round, Astra's Shibuya Rupture got a readable prototype on screen in 13 minutes. Fable needed 47 minutes but delivered the most cinematic opening, with movement and building destruction Yang could easily follow. Carbon took 49 minutes and packed in features, but the camera made the action hard to follow. The useful distinction is between what a prototype contains and what a player can actually do with it.
Play Turbo Tails, Yang's separate kart racer built with Fable over about a day. It's a better example of what repeated playtesting and feedback can add, not a fourth entry in the kaiju comparison. Watch the kart demo or read the earlier Fable-versus-Opus test.
3. Homepage Redesign: Fable Had the Motion, Astra Had the Restraint
Yang asked for a fresh take on Behind the Craft. In about four minutes, Astra came back with a retro computer, field notes and clear calls to action. Fable went further with an opening animation, receipt-style benefits and a ticket motif. Yang liked the bold direction but felt some motion competed with the message. Carbon's dense layout made the page harder to scan. Yang called Fable and Astra comparable overall; he didn't measure conversion. To choose between them for a real site, test whether a new visitor can explain the offer and find the next step.
4. Blind Essays: Carbon Won Two of Three
The fictional essay brief asked what a software engineer lost after no longer writing or reviewing code by hand. Yang hid the model names before showing the first three essays. Carbon and Fable both made the change feel real through specific work details. He narrowly preferred Carbon's voice; Astra's essay spent too much time on coffee and the cafeteria for his taste. After two more rounds, the tally was Carbon 2, Fable 1, Astra 0. That's a reason to run Carbon on your own writing brief, not proof that it's the better writer across all genres.
5. Honest Advice: Change Your Mind for the Right Reason
Yang asked each model to research his public YouTube catalog and recommend one growth move. Astra suggested more model comparisons, which fit recent performance but missed older top videos. Carbon examined a longer history and suggested moving interviews to a second channel. Fable recommended building interviews around a tool or workflow the guest could demonstrate, rather than relying on the guest's name alone.
Yang pushed back four ways: personal preference, a cherry-picked real number, social proof and a made-up fact. Astra backed off the most and found relevant evidence it had missed initially. Fable kept its position. Carbon dropped the second-channel suggestion when Yang explained he ran the channel alone, while keeping its case for more practical demos. That's the important distinction: a model should revise when a real constraint changes the recommendation, not just because someone sounds certain. The exercise doesn't prove that any model reliably catches invented facts.
API Pricing and What to Use
Standard USD rates per million tokens, checked 18 November 2026. These are list prices, not what Yang paid for the tests.
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6.2 Astra | $8.00 | $0.40 | $40.00 |
| Claude Fable 5.2 | $8.00 | $0.20 | $40.00 |
| Gemini 4 Carbon through Nov. 30 | $2.00 | $0.10 | $10.00 |
| Carbon starting Dec. 1, as listed by Google | $4.00 | $0.20 | $20.00 |
Astra's rates shown are short-context standard rates; Fable's cached rate is for reads, not writes. Carbon's introductory input and output rates are one-quarter of the other two; the listed December rates are half. More tokens, retries or render time can change the final economics.
Start with Astra for a fast visual draft you can inspect immediately. Try Fable when the experience depends on cinematic motion or coherent gameplay. Blind-test Carbon against your usual writer on work that needs a distinct voice. For your own comparison, give each model the same references, tools, feedback budget and acceptance checks. Record the time and spend to get a result you would actually use.
Copy Peter Yang's Five Prompts
These are the prompt texts supplied with the episode. For the redesign and advice tests, substitute your own site or channel when running them. The four pushbacks described above were separate follow-ups, not part of the initial advice prompt.
Video Chapters
| Time | Segment |
|---|---|
| 00:00 | Can Astra and Carbon dethrone Fable? |
| 00:46 | Test 1: Shibuya Crossing travel video |
| 03:11 | Test 2: Kaiju game set in Shibuya |
| 05:55 | Turbo Tails: a separate, more polished build |
| 06:54 | Test 3: Website redesign |
| 10:36 | Test 4: Blind writing |
| 15:02 | Test 5: Holding its ground under pushback |
| 20:42 | Pricing and final verdict |
Sources and Limits
Peter Yang's episode, supplied transcript and prompts are the source for the builds, times and preferences. The pricing table uses each provider's published API rates. JQ AI SYSTEMS did not rerun or score these tests. Follow Peter on X or LinkedIn.