Same Prompt, Four Models Turn In Results: What Accounts for the Difference Between $2.07 and $209.50?
Tony ran GPT-6.2 Astra, Sol, Fable 5.2, and Opus 5.5 on the same Alpine Gold prompt. The cheapest run was $2.07 and the most expensive was $209.50; the fastest was 30m 47s and the slowest was 3h 02m 54s. The video and bill are included.
Tony (@EnvolDev) posted a same-prompt comparison. With the same Alpine Gold prompt, four models each ran once, generating an airplane scene. Costs are estimated from API pricing, and timings come from run logs.
| Model | Cost | Time |
|---|---|---|
| GPT-6.2 Astra | $9.62 | 46m 14s |
| GPT-6.2 Sol | $2.07 | 30m 47s |
| Fable 5.2 | $209.50 | 3h 02m 54s |
| Opus 5.5 | $31.85 | 2h 45m 42s |
Tony's comparison video is below; it shows the airplane scene from this Alpine Gold run.
First, look at the price gap. Sol cost $2.07 this time; Fable 5.2 cost $209.50, roughly a 101x difference. Astra cost $9.62, about 21.8x cheaper than Fable, while cutting time from 3h 02m 54s to 46m 14s. Sol was about 15.4x cheaper than Opus 5.5 and about 5.4x faster. Tony also added in a reply that Sol was over 15x less money, over 5x faster than Opus. The numbers are right there; Fable's bill is hard to explain away with cinematic feel alone.
On quality, there is no single score. Someone in the comments mentioned that Fable 5.2 still had the most cinematic feel; Tony agreed and also said Astra's airplane detail was unreal. Another replier said Opus 5.5 was still the first choice for real work; Tony responded that Opus was still a solid pick, even with Fable 5.2 out. A single same-prompt run can show cost and speed, but it cannot directly pick a model for everyone.
The test boundaries should also be clear. Each model used the same prompt and ran once; costs are API-equivalent estimates; this comparison only includes OpenAI and Anthropic, not Gemini 4 Carbon. When someone asked where Carbon was, Tony said this one was OpenAI vs Anthropic; in an earlier Alpine Gold run, Gemini 4 Argon cost $11.61 and took 2h 01m 35s. That data is included here so readers can compare for themselves.
In a previous long post, Tony offered a rule: when comparing models, look at the cost to complete the task, not just the per-token price. These results put that rule on display again. Astra's per-token price is not low, but the total bill for completing the same job is far below Fable's; Sol is even more extreme, finishing for just over $2. Expensive models can still be worth it, provided that the quality gain lands exactly on what you need to deliver. In this airplane scene, Fable did not save the time back, yet its bill was an order of magnitude higher.
If you plan to reproduce this, don't treat a single run as final. In his long post, Tony recommends freezing the prompt, using an empty folder each time, recording tool versions and effort, running at least three times, and doing blind evaluation. There is only one run here, so it is suitable for seeing direction and price differences, not for giving models a final ranking.
Original post: https://x.com/EnvolDev/status/2122803967272866254
Long post: https://x.com/i/article/2120246086157881344
发布时间: 2026-11-18 09:29