Space Bunny REVEALED! MiniMax M3.1 Flash vs DeepSeek V4.1 Flash — Same Price, WILD Results?!

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Space Bunny REVEALED! MiniMax M3.1 Flash vs DeepSeek V4.1 Flash — Same Price, WILD Results?! · Channel: Prompt Engineer 48 · 09:55 · Watch on YouTube

Topics: AI evaluation and benchmarks, Software engineering with AI, AI models

Summary

The video identifies the anonymous OpenRouter model Space Bunny Alpha as MiniMax M3.1 Flash and compares it with DeepSeek V4.1 Flash, which costs the same at off-peak rates. Across four coding and front-end tasks, both models pass the bug-fixing, dashboard, and Tetris tests, with speed varying by task and hosting setup. DeepSeek is the only model to deliver a working 3D sneaker page; M3.1’s attempts either fail on small coding errors or return no answer. The creator stresses that these are limited, single-run tests rather than a definitive benchmark, and recommends checking which model is faster or cheaper on your own setup while keeping DeepSeek in the loop for complex builds.

Key points

OpenRouter’s Space Bunny Alpha was revealed as MiniMax M3.1 Flash.
M3.1 Flash and DeepSeek V4.1 Flash have the same listed price at DeepSeek’s off-peak rates: 15 cents per million input tokens and 60 cents per million output tokens.
MiniMax’s own benchmark table shows mixed results: DeepSeek leads on DeepSWE and Terminal Bench, while M3.1 leads on Automation Bench.
In the Python bug-fixing test, both models fix all eight bugs and pass all 18 tests; DeepSeek finishes faster.
Both models produce working revenue dashboards and Tetris games, with no clear overall winner across those tests.
DeepSeek delivers a working 3D sneaker page, while M3.1’s attempts fail due to code errors or return no answer.
The creator warns that speed results reflect two different hosting setups, not a pure model comparison.
Across the four runs, list-price costs were about 9 cents for M3.1 and 7 cents for DeepSeek; actual OpenRouter billing for DeepSeek was about 10 cents.
The comparison uses one run per task in most cases and does not replace independent, repeated benchmarking.

Chapters

Start	Chapter	Summary
00:00	Space Bunny revealed and the comparison setup	The creator identifies Space Bunny Alpha as MiniMax M3.1 Flash, compares its pricing and published benchmarks with DeepSeek, and explains the test setup and its limitations.
03:09	Python bug-fixing test	Both models fix eight planted bugs and pass all 18 tests, while DeepSeek finishes faster.
04:10	Dashboard and Tetris tests	Both models build functional dashboards and Tetris games, with speed advantages varying by task and no decisive overall winner.
06:43	3D sneaker page and final verdict	DeepSeek produces the only working 3D page, and the creator reviews the costs, caveats, and practical takeaway from the comparison.

Quotes

Now, OpenRouter’s own page has a banner developed and operated by MiniMax — 00:00

Both of these models think before they answer, and thinking tokens count against the cap. — 03:09

It’s an honest tie, and a useful one for this kind of front-end job, price and speed matter more than a leaderboard. — 05:12

on this job, DeepSeek is the only one that shipped a working page. — 08:22

It doesn’t replace a proper benchmark, and Minimax’s own numbers still need independent checks. — 09:24

Concepts and entities

MiniMax M3.1 Flash, DeepSeek V4.1 Flash, Space Bunny Alpha, OpenRouter, Ollama, Higgsfield, Terminal-Bench, AutomationBench

Full transcript (original language)

00:00 For about 4 weeks, an anonymous model called Space Bunny Alpha sat on OpenRouter free 1 million token contacts, and a lot of people were using it for coding. Nobody said who built it. Now, OpenRouter’s own page has a banner developed and operated by MiniMax, revealed to be MiniMax M3.1 Flash, and MiniMax’s launch post dated October 19th says the same thing. Here’s what I want to know. M3.1 Flash costs 15 cents per million input tokens and 60 cents per million output. DeepSeek V4.1 Flash

00:31 costs exactly that at off-peak rates. Same price. So, which one builds things that actually work? I gave both models four real jobs at the same moment with the same prompts, and I recorded every run. >> [music] >> Quick plan. First, what MiniMax claims, then four tests: a Python bug hunt, a dashboard Tetris, and a 3D sneaker page. The launch post says M3.1 Flash is nearly three times faster online than its predecessor. With usage cost for

01:02 agent tasks down 70% and a median output speed over 170 tokens a second. Those are MiniMax’s numbers. I looked for independent benchmarks this morning and found none for this model yet, so keep that in mind. What I like is that MiniMax put DeepSeek in its own table. That’s the newer DeepSeek V4.1 Flash, the 1017 snapshot. On DeepSWE, the coding benchmark, DeepSeek gets 77 and M3.1 gets 73.7. On Terminal Bench, 42.3 against 38.1. Where

01:35 M3.1 wins is Automation Bench, 62.8 against 60.5. So, even on MiniMax’s own table, this is not a clean sweep. One footnote matters. In the cost chart, DeepSeek is priced at peak rates, 30 cents in, a dollar 20 out. DeepSeek’s own pricing page says off-peak is 15 cents and 60 cents, the same as Minimax and peak is only certain weekday hours. If I have Deep Seek’s cost on their chart, which is my arithmetic and not theirs, the two land roughly level. So,

02:07 the price gap depends on when you call the API. Back to Space Bunny, OpenRouter’s chart shows about 4 trillion prompt tokens through it in under 2 weeks, but only 135 billion tokens of output. That’s roughly 30 tokens read for every token written. That’s what coding agents do, read a lot, write a little. Last thing, no open weights yet. This is API only for now, and Minimax says a bigger pro model is coming with no date. How I tested both models get the same prompt at the same moment on a page I

02:38 built. So, the clock you see is a real clock. M3.1 flash runs through Ollama’s cloud, which is free until October 22nd, 11:59 p.m. UTC. Minimax also offers it through its M plan and says the Minimax code app is included in the plan. I’m showing their pages, but I haven’t tested that app myself. Deep Seek runs through OpenRouter. That’s two different roads, so read the speed numbers as how this setup behaved today, not as a pure model race. For cost, I use the official list price for both, and I’ll tell you

03:09 what OpenRouter actually built. One thing burnt me, so let me save you the trouble. My first attempt capped the output at 32,000 tokens. Both of these models think before they answer, and thinking tokens count against the cap. Four of my six runs ended with no finished file, and in one of them M3.1 spent the entire budget just thinking. So, I raised the cap to 100,000 and reran everything you’re about to see. One try per task except one, and I’ll flag it. Test one, a small Python project with eight bugs planted across two files, a

03:39 date range function that skips the last day, a cache that never refreshes what you just used, a default list shared between objects, and off-by-one in pagination. 11 of 18 tests failed. Both models get the code, the tests, and the failure output. Go. DeepSeek finished in 13 seconds. M3.1 took 35. Both fixed all eight bugs and 18 out of 18 tests passed. So, on

04:10 correctness, it’s a tie. The difference is the thinking before the answer. M3.1 produced about 4,000 tokens, DeepSeek under 2,000. At these prices, that’s a fraction of a cent either way. So, for a small fix, the real cost is your waiting time, and here DeepSeek was faster. Different roads, though, so I’d call that a hint, not a verdict. Test two, a revenue dashboard for a made-up coffee roaster. Four KPI cards, a line chart, a stacked bar chart, and a donut, all drawn by hand with no chart library. Region checkboxes that update

04:40 everything, tooltips, dark theme. DeepSeek was a bit quicker, 184 seconds against 241, but it wrote more tokens, about 43,000 against 37,000. So, at list price, it cost slightly more. Both work. I check in Chrome, no console errors, and unticking a region changes the charts and the totals.

05:12 I also checked the growth math on both KPI cards, and it’s right on both. M3.1 gives each region a legend chip with its yearly total. DeepSeek goes for a cleaner area chart and a quieter layout. Neither scrolls sideways at phone width. If you gave me these two screenshots without names, I couldn’t tell you which is the better product. It’s an honest tie, and a useful one for this kind of front-end job, price and speed matter more than a leaderboard. Test three, Tetris. This one is a good trap because there’s a lot to get right

05:42 at once. Seven pieces, rotation near a a hole slot, a ghost piece showing where it lands, a next piece preview, and speed that rises with level. Here, the clocks flip, M3.1 finished in 3:13, Deep Seek took 4:12, but look at the tokens, about 21,000 against 45,000. Deep Seek’s host was streaming at only about 50 tokens a second this time, which is a hosting detail on Open Router, not the model. That’s exactly why I don’t trust a single speed number.

06:13 Both play properly, pieces fall, the ghost lands where it should, hold works, and the score counts up. M3.1 has a pause button and a speed panel. Deep Seek has a tighter board, my checks only catch crashes and basic behavior, so I can’t tell you whose rotation feels better. If you want a winner on Tetris, there isn’t one. Both passed.

06:43 Test four is the one I was most curious about, 3D product pages. To be clear about who does what, Minimax doesn’t make 3D models. I used Higgsfield to generate a generic sneaker with no logos, then turn that picture into a textured 3D model that cost about 10 and 1/2 credits in total. The model file is 45 megabytes. The AI models only write the code that shows it 3JS, a camera that moves as you scroll, auto rotate until you drag, color swatches that recolor the shoe, and a loading indicator.

07:18 Deep Seek’s page works from the first load. The hero animates, the camera moves through the sections. As I scroll, I can drag the shoe, and the swatches change the colors. No console errors and no sideways scroll at phone width. M3.1’s first page did not work. It loads, then shows the model could not be prepared. The console says assignment to constant variable. Here’s why. On line

07:50 1514, it declares model radius with const on line 1681. It assigns to it again. One keyword, const, instead of let, and the whole page stops. Then I ran M3.1 again on Ollama twice. The model thought 4 minutes and returned no answer at all. So, I tried the same model through OpenRouter to separate the model from the hosting. That one did deliver a page, and it broke, too. Line 1960 has a typo, glltf, a variable that doesn’t exist. The page

08:22 says unable to load. So, that’s two pages out of two stopped by a single small mistake, plus two empty answers. In fairness, DeepSeek also got only one finished try here, so I’m not saying it always works. But, on this job, DeepSeek is the only one that shipped a working page. Scoreboard: bug fixing, a tie on correctness, DeepSeek faster. Dashboard, a tie, DeepSeek faster. Tetris, a tie, M3.1 faster. The 3D page, DeepSeek worked, M3.1 broke twice on one-word

08:53 bugs and returned nothing twice. At list price, all four runs together came to about 9 cents for M3.1 and 7 cents for DeepSeek. On OpenRouter, DeepSeek actually billed me about 10 cents because one host charged double the list price on the dashboard run. When I checked hosts listed that same DeepSeek model from about 6 cents to 60 cents per million input tokens, the top one being a fast tier. M3.1 cost me nothing on Ollama, but that’s only until October 22nd. What this does not prove, it’s one

09:24 run per task, my prompts, two different hosting setups, and a free window for one of them. It tells you how these models behaved for me this week. It doesn’t replace a proper benchmark, and Minimax’s own numbers still need independent checks. So, here’s my read. If your job is quick fixes and front-end pages, either one works and whichever is free or faster on your setup wins. For long tricky builds like the 3D page, I’d keep Deep Seek in the loop until M3.1 shows it can finish that kind of job reliably. I’ll rerun

09:55 these a few times once more people have tested it and on Minimax’s own M plan road, if you’re looking for cheap and reliable GPUs, I use RunPod and the link is in the description. Tell me which test you want next and what you thought of these two. See you in the next one.