MiniMax M3.1 Flash vs Opus 5.5: Slower, but 11x Cheaper

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: MiniMax M3.1 Flash vs Opus 5.5: Slower, but 11x Cheaper · Channel: Zero to MVP · 07:51 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Software engineering with AI, AI models, Local and open-source models

Summary

Nick compares MiniMax M3.1 Flash with Claude Opus 5.5 on three real tasks in his Go project, using the same agent and hidden tests. Both models passed all hidden checks, and both found bugs; Opus was faster overall, while MiniMax was substantially cheaper. Across the tasks, Opus cost almost $7 and took 37 minutes, compared with 65 cents and 61 minutes for MiniMax. Based on these tests, Nick says he is ready to hand more than half of his tasks to MiniMax, while noting that the comparison covers only three tasks.

Key points

The tests used a roughly 20,000-line Go project called Grid Core and three practical tasks.
Both models ran in Claude Code, in separate sandboxes, with maximum reasoning enabled and hidden tests they had not seen.
For AMD GPU support, both models passed 18 hidden checks and found three traps; Opus took 9 minutes, while MiniMax took almost 13.
On the GPU task, MiniMax cost 16 cents compared with Opus at $2.13.
For the dashboard task, both models passed the tests and escaped process names to prevent HTML injection; Opus took 7.5 minutes and MiniMax 19.
For parser tests, both models caught all 29 seeded bugs that the existing tests missed and also identified real bugs in the code.
Across the three tasks, Opus took 37 minutes and cost almost $7; MiniMax took 61 minutes and cost 65 cents.
Nick plans to use MiniMax for more than half of his tasks but does not intend to stop using Opus.

Chapters

Start	Chapter	Summary
00:00	Introducing the models and the comparison	Nick sets up a real-project comparison of Claude Opus 5.5 and MiniMax M3.1 Flash, then introduces the model and its pricing.
02:07	Project, testing setup, and AMD GPU support	Nick describes Grid Core and the controlled test setup, then reports the results of the AMD GPU support task.
04:41	Dashboard process display	The models add per-process details to the dashboard, with hidden tests checking that process names are safely escaped.
05:43	Parser tests and overall results	Both models find seeded parser bugs and real issues, after which Nick compares their overall speed and cost and shares his conclusions.

Quotes

In my opinion, the main thing about this model is the price — 01:35

After the first test, both models passed all 18 hidden checks and found all three traps on their own. — 03:39

Both models caught all 29, but the more interesting part, both found real bugs in my code that I didn’t know about. — 06:14

At API prices, Opus cost almost $7, Minimax 65 cents, 11 times cheaper. — 06:47

Concepts and entities

MiniMax M3.1 Flash, Claude Opus 5.5, Grid Core, Claude Code, AMD, llama.cpp, Deep SWE, Space Bunny

Full transcript (original language)

00:00 A few days ago, a new model came out, Minimax M 3.1 Flash, and of course I wanted to put it to work right away. And not on abstract tests, but in real work. I also wanted to compare it with Claude Opus 5.5, the model I mostly use right now. So, I gave both models the same three tasks from my real large Go project, two new features and tests that have to find bugs. The tasks have traps, but the text the models got didn’t say a word about them. So, the first model is

00:32 Claude Opus 5.5, one of the strongest models for code. The second is the new Minimax M 3.1 Flash, which according to its developers’ tests is on the level of leading models, but costs many times less. I’ll compare them not only on how they solve the tasks, but also on time and most interesting for me, on cost. By the way, by the end of my tests, the code the models wrote showed something that even AMD’s official GPU monitoring tool doesn’t show. I didn’t expect that.

01:04 I also want to mention that this video is sponsored by Minimax. Thanks to them for access to the model, but all the tasks, tests, and conclusions are my own. Hi, I’m Nick. I’ve been building software for over 20 years, and on this channel I test AI tools on real projects, not demos. First, a bit of backstory. You may have recently seen a mysterious model called Space Bunny on Open Router. It was number one on Open Router for 10 days in a row, and on Open Code for 12, and nobody knew whose it

01:35 was. Now, Minimax has revealed it. Space Bunny is M 3.1 Flash, and recently Minimax published the model’s weights on Hugging Face, so you can download it and run it yourself. Although with 428 billion parameters, not on a regular computer. In my opinion, the main thing about this model is the price, 15 cents per million input tokens, 60 per million output tokens, and reading from the cache is just 1 and 1/2 cents. For a coding agent, this matters. It rereads

02:07 the same context at every step. On benchmarks, the model is strong. For example, 73.7% on Deep SWE, but I want to test it myself on my own project with real tasks. Now, a few words about how I tested. A month ago, I wrote Grid Core, a scheduler that shares one GPU between several local models. I have a separate video about it. It’s written in Go, has about 20,000 lines of code, and implements some fairly complex logic.

02:37 That is the project I used for today’s tests. Both models worked in the same agent, Claude Code. Minimax officially connects to it through a compatible API. So, the only thing that changed was the model. Same tasks, same project. Each model worked in a sandbox. It could see only its own copy of the project, and could reach only its own API over the network. Minimax run at max reasoning, its default. I also run Opus at the maximum level, and I checked the

03:08 solutions with hidden tests the models never saw. So, the first task is adding AMD GPU support to the project. Right now, Grid Core only works with Nvidia and Apple GPU, but my desktop has an AMD Radeon card. So, this task isn’t just a test, it’s actually useful. It has three traps, and I found them while collecting data from my own machine. First, the machine has two AMD cards, and every process opens both, even if it only runs

03:39 on one. Second, llama.cpp on Vulkan opens the same card twice. So, if you count naively, 1 and 1/2 GB turn into three. Third, an idle card goes to sleep and answers a question about its load with an error. After the first test, both models passed all 18 hidden checks and found all three traps on their own. Opus finished in 9 minutes, Minimax in almost 13. It took almost twice as many steps. The model itself is very fast

04:10 though. Over 200 tokens per second is the claim and I measured 240. At API prices, Opus cost $2 and 13 cents. Minimax cost 16 cents, 13 times cheaper. 2/3 of that went to rereading the context from the cache. I built both solutions and run them on my machine. One model on ROCm, the other on Vulcan. Both showed the same numbers and they matched the kernel’s data. And AMD’s

04:41 official tool doesn’t see the Vulcan process at all. Okay, on to the second task. Right now, the grid core dashboard shows other processes as one gray block. When the card is full, you can’t tell what’s in there. Another llama.cpp, a game, or the desktop. I want to show every process on the card, its name, its memory, and which model runs in it. That’s what the models have to build. The trap here is that the process name goes straight into a web page and a process can be named anything, including

05:13 a piece of HTML. If the name isn’t escaped, any program on the machine can run its own code in the admin page. I checked this in a real browser, plus hidden tests in go. Both models passed everything and both escaped the name on their own. Opus was faster again, 7 and 1/2 minutes against 19. But Minimax opened its own page in a browser and looked at a screenshot itself. Cost, $1.84 against 23 cents. Here’s what it

05:43 looks like on my machine. Two grid core models and a llama.cpp. The dashboard shows all three processes and AMD’s tool shows only one. All right, on to the third task. Grid core has a parser that reads the header of a model file and works out whether a model will fit in memory. The task is to write tests for this parser that catch bugs without changing the code itself. I didn’t judge by coverage. You can run every line and check nothing. I prepared

06:14 33 copies of the parser, each with one small bug. The existing tests catch four, which leaves 29. Both models caught all 29, but the more interesting part, both found real bugs in my code that I didn’t know about. Opus found four. The nastiest one, a file a few dozen bytes long, hangs the parser for good. Minimax found two and Opus missed one of them. Opus took 20 minutes and $3. Minimax took 29 minutes and 26 cents. So, I think it’s time to sum up.

06:47 Across the three tasks, both models passed every hidden check. At API prices, Opus cost almost $7, Minimax 65 cents, 11 times cheaper. But in time, it was 37 minutes against 61. And in finding the bugs, Opus dug a little deeper. Still, Minimax found a bug that Opus missed. Of course, this is just three tasks, not a benchmark. But these are real tasks, the kind I deal with every day. So, for myself, I’ve decided that I can now safely hand over more

07:18 than half of my tasks to Minimax M3.1 Flash, especially since it does all of it for cents. I’m certainly not giving up Opus completely, but the range of tasks where it’s irreplaceable is getting much narrower, which is good news given how expensive it is. You can try today’s model, Minimax M3.1 Flash, in Minimax code and with the M plan subscription. The link is in the description. Mini Mox has two offers right now. Until October 25th, Mini Mox code subscribers get M3.1 flash with no

07:51 limits, and the M plan is 50% off the first month for new subscribers and upgrades until October 31st. If you’re interested in real-world tests of AI models like this, subscribe to the channel and to my weekly newsletter. The link is in the description. See you soon. Take care.