The Engine Finished in 9 Minutes: Why Some Would Rather Have the 2-Hour Version
Four models generated cutaway views of a four-stroke engine from the same prompt: the fastest took 9m 09s, the cheapest cost $0.36, and the most expensive cost $21.81. Put speed, detail, and cost side by side, and the ranking matters much less.
With the same prompt, four models turned in their work: a cutaway view of a four-stroke engine. Prices ranged from $0.36 to $21.81, and times from 9m 09s to 2h 00m 13s.
| Model | Cost | Time |
|---|---|---|
| Claude Opus 5.5 | $21.81 | 2h 00m 13s |
| GPT-6.1 Terra | $0.43 | 9m 09s |
| GPT-6 Astra | $7.33 | 54m 33s |
| DeepSeek V4.1 Flash 1017 | $0.36 | 1h 54m 16s |
The video comes from the original tweet and shows a comparison of the four-stroke engine cutaway views generated by the four models from the same prompt.
Original poster Tony's conclusion is a little counterintuitive: it's no surprise that Opus got it right; DeepSeek is as good as Claude, and may even be the most beautiful of the four, for only 36 cents—about 60 times cheaper than Opus. Terra is fast, but it costs more than DeepSeek, which he finds amusing.
A bit of context here. A four-stroke engine cutaway is not just about drawing an outer shell; the relationships among the cylinder, piston, connecting rod, crankshaft, valves, and timing need to line up. It's not hard for a model to make the exterior look good; the hard part is keeping the structure from contradicting itself.
The comment section immediately split into two camps. OwenZ asked: for roughly the same 2 hours, with a 60-fold price difference, where exactly does DeepSeek go wrong? Tony replied that he liked DeepSeek's result best—it did a lot of work for $0.36. Arbaz's challenge was more direct: Terra took 9 minutes, yet you crown the 2-hour version? Tony's take was that Terra wins on speed, DeepSeek wins on detail and completeness, and he was weighing quality, time, and cost.
Another reply raised the timing issue. CryptoBandit said 1 hour 54 minutes was the number he would look at. He ran the same model on his own four GB10 machines, EXL3 4.0 bpw, and in 14 minutes 44 seconds wrote a runnable 3D road-and-river crossing game: 80,935 tokens, successful on the first try. No local queue. Tony responded that his timing includes execution and rendering, and that only same-task comparisons make things clear.
This debate is more useful than who came in first. Put speed, price, and completeness in one table, and the answer depends on what you're using the model for. If you need a fast draft, Terra's 9 minutes is hard to ignore. If you want to control costs and are willing to wait, DeepSeek's 36 cents stands out. If you need reliable delivery and aren't budget-sensitive, Opus is still Tony's daily workhorse. In the end, he also admitted that DeepSeek had caught his attention this time.
Original tweet link: https://x.com/EnvolDev/status/2116189885723734309
Share link: https://x.com/LeeLeepenkman/status/2116304331268108544
发布时间: 2026-10-31 07:30