Mercury 2.5 vs Solar Pro 4: is a diffusion LLM really faster?
Yes. Mercury 2.5 answered 7.3x faster than Solar Pro 4 on average, and was faster on 40 of 43 prompts. What it does with that speed is the more interesting part: it writes about half as much, and still uses more output tokens.
Mercury 2.5 is a diffusion LLM (dLLM). Instead of generating one token at a time, it drafts many tokens in parallel and refines them, which is what Inception points to as the source of its speed. We put it against Solar Pro 4, a conventional model in the same low price band, on the same 43 prompts.
How we measured
Both models got the 43 prompts from FastMetal's comparison set under identical conditions, with no reasoning setting specified. Mercury 2.5 thinks at medium effort by default, so this is its out-of-the-box behavior. The run was in September 2026.
Results
| Mercury 2.5 | Solar Pro 4 | |
|---|---|---|
| Mean response time | 3.8 s | 27.7 s |
| Mean output tokens | 1,968 | 1,088 |
| Mean answer length (characters) | 1,457 | 3,296 |
Mercury 2.5 was faster on 40 of 43 prompts: 7.3x on average and 26x on the prompt with the widest gap. That isn't a difference you notice; it's an order of magnitude.
What the speed is made of
Mercury 2.5 gave the shorter answer on 42 of 43 prompts, about half the text of Solar Pro 4. Yet it used about 1.8x the output tokens.
Half the readable text and 1.8x the billed tokens: the difference is thinking. With reasoning on by default, Mercury spends tokens behind every short answer. The effect is strongest on short prompts, where a fixed thinking overhead is a larger share of a small answer.
What it costs per request
The prices have moved since this run, and in Mercury's favor. Today Mercury 2.5 is $0.044 per million input tokens and $0.165 per million output; Solar Pro 4 is $0.099 and $0.396. At those rates the mean output cost per 1,000 requests is about $0.32 for Mercury 2.5 and $0.43 for Solar Pro 4, and Mercury's input rate is less than half of Solar's. So even with 1.8x the tokens, Mercury 2.5 now comes out cheaper per request on this prompt set. (When the two were priced closer together, the extra tokens made Mercury the more expensive one, so this depends on the rate card, not the model.)
Cheaper per request isn't the same as cheaper per unit of answer, though. Solar Pro 4 writes twice as much for its 30% higher output cost.
Which one to use
- Choose Mercury 2.5 when response time shapes the experience: chat UIs, completions, anything that has to answer on the spot. A 7x gap is hard to make up any other way, and its habit of short answers suits those uses.
- Choose Solar Pro 4 when you want long, detailed answers and can wait: batch jobs, drafts, explanations. You get about twice the text per request.
You can also cut Mercury's thinking. Set reasoning to {"enabled": false} or {"effort": "low"} and the token count drops. This comparison used the defaults only; how to set it is in reasoning control.
The full answers are side by side on the comparison page. Details are on the Mercury 2.5 and Solar Pro 4 pages, and current prices on pricing. The API model ids are mercury-2.5 and solar-pro4.