When you ask He.art a question, you can choose which model answers it. The picker offers fourteen models from six companies: MiniMax, Google, Anthropic, OpenAI, xAI and DeepSeek. That is a lot of choice, and the names alone tell you little about what you are choosing. Which one starts writing first? Which one costs the least? Does the expensive one do better on an everyday question?
So this morning we asked them. Every model got the same three questions, twice, and we wrote down how long each answer took, how many tokens it used and what it cost. This post is what we found, including a few things about He.art itself that we did not expect.
How we tested
We kept it small and simple on purpose, so that anyone can read every answer and check every number.
- The same code He.art uses. Each question went through He.art's own provider code, the functions that answer you in He.art chat, with the same instructions He.art gives every model for a short answer (He.art answers short first, and gives a longer answer when you ask for one).
- Web search off. In He.art chat, search is on wherever a model supports it. For this test we turned it off, so that no model got outside help and no search fees muddied the costs.
- Three questions, twice each. An explanation, a small math problem and a small coding task. That makes six answers per model and 84 in all, asked on Oct 21, 2026 between 07:22 and 07:28 UTC.
- Timed from the outside. We timed two things: how long until the first words appeared, and how long until the answer was complete. The requests went from a PC on a home internet connection, not from He.art's servers, so every time includes an ordinary network trip.
- Priced from the price lists. Token counts are the ones each provider reported with its answer. Costs use each company's published price that day: Google, Anthropic, OpenAI, DeepSeek and MiniMax. For xAI we used the cost its API reports with every answer, which matches the prices on its models page.
The three questions, exactly as asked:
- "Why is the sky blue? Explain it for a curious ten-year-old."
- "A bakery sells muffins at $3 each or 4 for $10. Maya needs 11 muffins. What is the least she can pay, and how should she buy them?"
- "Write a JavaScript function that takes two dates as strings in the form YYYY-MM-DD and returns the number of whole days between them."
Every answer, with its timings, tokens and cost, is in the raw results.
Everyone got it right
We expected at least one stumble. There was none.
All fourteen models answered the muffin question correctly both times: two packs of four ($20) and three single muffins ($9), for $29. It is a small trap, because the cheapest-looking move (three packs, twelve muffins, $30) costs a dollar more.
For the coding task, we took the function out of each answer and ran it against six pairs of dates: two months that include a leap-year February, a single month, the same day twice, a span across a leap day, a span across New Year, and a pair given in reverse order (we accepted either a positive or a negative count for that one). All 28 functions passed all six.
The explanations of the blue sky were all accurate too. They differed in style rather than substance: some named Rayleigh scattering, several added a word about red sunsets, and a few reached for emoji.
On questions like these, then, the choice of model is not about getting the right answer. It is about how long you wait and what it costs.
Speed and cost
Here is every model, with the median of its six answers. "First words" is how long until text began to appear; "Whole answer" is how long until it was done. The cost is for 1,000 questions like ours, so the small numbers are easier to compare.
| Model | First words | Whole answer | Tokens out per answer | Cost per 1,000 questions |
|---|---|---|---|---|
| MiniMax M3 (He.art's default) | 1.4 s | 3.8 s | 163 | $0.42 |
| MiniMax M2.7 Fast | 3.3 s | 5.8 s | 386 | no published price |
| Gemini 3.5 Flash | 3.5 s | 3.8 s | 633 | $2.43 |
| Gemini 3.1 Pro | 5.6 s | 6.4 s | 620 | $7.58 |
| Gemini Flash-Lite | 0.7 s | 1.1 s | 114 | $0.19 |
| Claude Opus 4.8 | 1.2 s | 3.2 s | 152 | $4.34 |
| Claude Sonnet 5 | 1.8 s | 3.9 s | 315 | $3.36 |
| Claude Haiku 4.5 | 0.6 s | 2.0 s | 138 | $0.77 |
| GPT-4o | 0.7 s | 1.5 s | 111 | $1.31 |
| GPT-4o mini | 1.0 s | 1.9 s | 114 | $0.08 |
| Grok 4.5 | 3.6 s | 4.1 s | 265 | $1.96 |
| Grok 4.3 | 5.7 s | 6.1 s | 481 | $1.40 |
| DeepSeek Chat | 0.7 s | 1.3 s | 101 | $0.14 |
| DeepSeek Reasoner | 1.6 s | 2.3 s | 268 | $0.35 |
A few things stand out.
The quick ones are very quick. Claude Haiku 4.5, Gemini Flash-Lite, GPT-4o and DeepSeek Chat all began writing in under a second, and Flash-Lite, DeepSeek Chat and GPT-4o had finished within a second and a half.
The cheap ones are very cheap. GPT-4o mini answered for 8 cents per thousand questions, DeepSeek Chat for 14 cents and Gemini Flash-Lite for 19 cents. At the other end, Gemini 3.1 Pro cost $7.58 for the same thousand questions, more than ninety times as much as GPT-4o mini, and Claude Opus 4.8 cost $4.34, for answers that were no more correct here.
Thinking is billed, even when you never see it. Gemini 3.5 Flash wrote the shortest answers of the Gemini family (a median of 60 words), yet it used 633 output tokens per answer. Most of those were spent thinking before it wrote: on the muffin question, one answer used 381 tokens of thought for 41 tokens of reply. Google's price list counts thinking tokens as output, so that thinking is most of what you pay for. Gemini 3.1 Pro, both Grok models, MiniMax M2.7 Fast and DeepSeek Reasoner think first too, and it shows in their times to the first word. Claude Sonnet 5 thought briefly on the muffin question (280 and 314 tokens) and hardly at all on the others. Claude Opus 4.8 reported no thinking on any of the six, which is why the most expensive Claude model was still one of the quicker ones to start.
Expensive is not the same as slow, or cheap as fast. GPT-4o mini, the cheapest, was a little slower to start than GPT-4o. Claude Opus 4.8, the priciest Claude model, started faster than Claude Sonnet 5.
One answer took 44 seconds. Grok 4.5's very first answer waited 41.7 seconds before its first word. Asked the same question again a few minutes later, it took 3.1 seconds in all. Six answers per model is enough to see a pattern, not enough to rule out a bad moment, which is why the table shows medians.
What we learned about He.art
A test like this checks the tester as well. Three things in our own setup surprised us, and you should know them.
"Gemini 3.5 Flash" is now answered by Gemini 3.6 Flash. Every answer Google sent back for the model He.art calls Gemini 3.5 Flash was labeled gemini-3.6-flash, and Google's price list no longer has a 3.5 Flash at all. We priced it as Gemini 3.6 Flash, $0.75 per million input tokens and $3.75 per million output tokens; Google lists those prices through Dec 31, 2026, and twice that from Jan 1, 2027. The name in our picker is out of date, and that is on our list too.
Both DeepSeek choices are one model. DeepSeek Chat and DeepSeek Reasoner were both answered by the model DeepSeek now calls deepseek-flash, the only Flash model on its price list. They still behave differently: Reasoner thinks before it answers and Chat does not. Our runs fell inside DeepSeek's weekday peak hours (which include 06:00 to 10:00 UTC), so they were priced at the peak rate, which is twice the off-peak one.
Every Grok question was asked twice. He.art asks xAI to search the web with each question, through a feature xAI has since retired: each first request came back refused with a note that "Live search is deprecated". He.art then asks again without it, so you still get your answer, but every Grok answer pays for a wasted round trip first, and Grok in He.art has not been searching the web. That one is ours to fix, and it is on our list.
One more gap: MiniMax M2.7 Fast does not appear on MiniMax's price list, so we could not price it.
What this does not tell you
Three questions are not a benchmark. These were short, everyday questions with clear answers, asked one morning from one connection. Longer writing, harder reasoning, images and documents could separate the models in ways this test cannot, and times on another day will differ. We also judged only whether each answer was right, not which one was nicest to read; you can do that yourself from the raw results.
The 78 answers we could price cost 15 cents in all.
Try it yourself
The model menu is inside the box where you type in He.art. Ask one of these questions, or one of your own, then switch models and ask again: the answer you prefer is the best test there is. We will be back next week with more.