Truly ambiguous use of screen space, enough overlapping windows to turn the desktop into a stress test for your eyes. That app and that roast came from MiniMax M3.1 Flash. A model has only just come out. Flash models are meant to be the cheap, slightly dim option you use when you don't want to pay for a proper model. MiniMax reckon this one codes like the top models for a fraction of the price. And I wanted to find out if that was true. MiniMax have just released M3.1 Flash. On paper, it's a big jump. They say it's nearly 3 times faster than the latest model and 70% cheaper to run. With benchmark scores sitting right next to the top models. Launch charts are easy to make look good though. So, I put it up against 5 of the best flash models out there. That includes Mercury 3, a diffusion model from Inception that launched on the same day and claims over 1000 tokens a second. I put them through a speed race, a business automation job, and a vision test. And then I had M3.1 build that roaster app you saw from scratch and fix a real bug in it. By the end, you'll know whether it lives up to the launch and which of these flash models deserves an agent budget. There's also one setting MiniMax switches on by default that changes a lot. In my speed test, it left M3.1 thinking for up to 47 seconds before it wrote a single word of its answer. I'll show you the 1 line fix near the end, so definitely stay tuned for that one. This video is sponsored by MiniMax, so thank you, MiniMax. But the tests and the numbers you're about to see are all mine. Let's get on with it. First up is speed. This is Flash Bench, a little tool I built for this video. And it fires up the same prompt at all 6 models at once. So, DeepSeek Flash coming straight in there with a really fast result, 6.42 seconds overall. Tool Mercury 3 coming in second. Hi Q coming in third. Across 11 races, DeepSeek won 7 Mercury won the other 4. M3.1 didn't win any and it typically came 4th or 5th. It was the 3rd cheapest every time behind Mercury and Haiku at around a 3rd of a cent to run at most, whereas Gemini cost up to 10 times as much. What is really interesting is how the speed claims held up. Inception say Mercury does 1068 tokens a second and it mostly ran at 330 and 430 with its best run hitting at 832. So, it's quick, but nowhere near that claim. And then Mercury 3 coming in last there didn't follow instructions and wasn't very correct. So, it really did prioritise speed over accuracy. Minimax claim 170+ for M3.1 and it mostly sat between 90 and 150 with 1 run getting up to 174. The coding prompt was the most telling. Mercury wrote the whole parser and its tests in under 4.50 seconds for 0.006 of a cent. M3.1 came in 5th in about 30 seconds and almost all of that was spent thinking before it wrote a single line. All 6 parsers passed their own tests. Though on this one, it really comes down to speed and cost and Minimax really wins on that cost. Thing that stood out to me though is how long M3.1 sits there thinking before it answers. On quick maths question, it was only a few seconds, but on the longer prompts, it was about 21 and 47 seconds before it wrote a word. So, hold that thought because we'll come back to it. Speed's nice to have, but it's whether the code works is the stuff that matters. Minimax put M3.1 at 73.70 on Deep Seek. That's a benchmark where the model has to understand a code base, build or fix something and then prove it works by running the tests. Deep Seek reports 77 for V4.1 Flash on the same benchmark. Rather than trust either chart, I gave M3.1 a real bug in the Roast app you saw at the start. Closing the app made Python crash and rather than guessing, M3.1 opened and closed it in different ways until it found the cause. The cause was the buttons destroying the window from inside its own click handler, which corrupts the memory TK uses and kills Python on the spot. It narrowed that down to a 10 line repo and then fixed it, and the app now closes cleanly. That's exactly the kind of work Deep Seek is meant to test, and honestly, if there's one thing that you take from this video, is that M3.1 is a coding beast. If that just saved you picking a model off a launch chart, leave a like. It's also worth you leaving down in the comments which flash model your agents run on, and roughly what you spend a week, so other people can see the real numbers. If enough of you want Deep Seek or GLM, I'll do a self hosted version next. Next up is automation. MiniMax says M3.1 is one of the top performers on automation bench, and it scored 76.20 on Banker Tool Bench, which tests the things like planning several steps ahead and keeping details straight across different documents. Rather than take their word for it, I tested it on something I deal with every single week, which is sponsorship emails. Each model got 10 anonymised emails from my inbox, a copy of my sponsors table, and 4 tools. It could read an email, search the sponsors, and update or add sponsor, and draft a reply. I also hit a trap in there. One of the emails comes from the same person at different address, so the right move is to update the existing entry rather than add a 3rd. Deep Seek is already searching for sponsors, GLM Flash is trailing behind, still reading those emails. So, it's going to do this and and work within the tool set that it's got and help me sort out my email. So, looking at the medians of all the results here, Deep Seek coming out on top, winning 2 of them, averaging at a place of 1.50, scoring 100% on all of those tests, so really, really good at agentic use. MiniMax coming in 3rd, averaging 3.50, I guess, passing 3 of the 4 tests. Gemini 3.80 trailing slightly behind, losing mostly on time. HiQ doing absolutely terribly, only passing 1 of the 4 tests in GLM 5.30 flash, completing 2 of the 4 tests. So, 2 very, very disappointing models there for Agentic work. Not 1 of the 6 fell for it, which I didn't expect. Mercury was the surprise of the round, passing all 4 runs while being the fastest and cheapest, under half a cent a run. Gemini passed every time, too, but at $0.18 a run, it cost over 20 times as much as M3.1. M3.1 got right 3 times out of 4, but it only said it had finished once. It was using 37 to 40 tool calls, and 40 is the limit I set. So, it was still checking its own work when it ran out of steps. When it did work, it really did match the best models. Running out of steps comes down to that same setting I mentioned at the start, which I promise we'll come back to. For vision, I went a step further and gave the models short video clips rather than a screenshot. Only 3 of the 6 can watch the video, which are M3.1, GLM, and Gemini. So, DeepSeek, Haiku, and Mercury sat this one out. I have a video, thumbs up to the camera, let's see what that does. Let's upload this. What is this video of? And is it a positive or negative sentiment? The video's a man at a desk turning towards camera, smiling, and giving a double thumbs up. The sentiment's clearly positive, conveying approval and enthusiasm. A new video I've just recorded. You really need to subscribe. What is the man saying? The gentleman says, 'You really need to subscribe.' I can't determine his exact words from still images, so again, this is the limitation there. So, both GLM 5.30 flash and MiniMax 3.1 flash, they're visual models without any audio, whereas Gemini 3.80 flash nailed it. All 3 got the sentiment, the hat, and the silent clip right. And M3.1 was the cheapest every single time. That made Gemini the only one to get all 4, and it was the fastest, too, at about 5 seconds an answer. On price, again, coming on top, costing a 3rd less than Gemini did, and 18 times less than GLM 5.30, which was the most expensive and no more accurate. If your agent needs to read what's on screen, M 3.1 is the cheapest one that gets it right. But, if the answer depends on sound, you want Gemini. Now, let's take a look at how that roster app was made. It took M 3.1 4 prompts and about 1.67 hours of working on its own. If we look at the code, it's all in 1 Python file, about 1600 lines, but it's split up sensibly into screen capture, the API calls, the providers, and the window. It really got the important stuff, too, which I expect from a coding model, like telling you exactly how to fix the Mac OS screen recording permission and falling back to a screenshot if FFmpeg isn't installed. It also wrote 53 unit tests, and they all passed. The build used 34m tokens, and 33m of those were cached reads, which is the model rereading the code it already seen at a fraction of the price. Without caching, the build would have cost about $5 instead of $0.77. To keep myself honest, I also had Claude Opus review the app and its code. Its verdict was that it's a single, well written Python file. The 53 tests passed, and the real risks aren't in the code quality, but in the privacy and in how it handles things going wrong. Now, what I want to dig into here is design, because I had M 3.1 make 2 landing pages for the roster. One used a design skill, which is a set of instructions telling the model how to approach design, and the other had no skill at all. This is a result of me just telling it to build a landing page and then direct it or anything like that, and mostly it looks okay, I guess. A few little spacing issues here. I'm not looking particularly at the actual copy itself, but no, it looks pretty good. Just a little bit dull. Now, interestingly, I then asked it to recreate another landing page, but don't use any design skills or anything like that. I did notice it used skills for this one, AI design skill, but this one was no design skill or no skills even. And I would say actually the raw output looks a lot better. Still a little bit AI generated, not particularly creative or interesting. However, it does I think objectively look better than this result. So, how much did all of this use cost then? Cuz everything I built for this video plus every test I ran, I used about 66m Minimax tokens. As wild as that sounds, the whole lot came to about $1.66 at list price. That's because 64m of those tokens were cash reads, which is the model rereading things it already seen at a 10th of the normal price. Without caching, the same work would have cost over $10. So, what does all this cost to use day to day? There are 4 ways in. The API is $0.15 per million tokens going in, $0.60 coming out, and $0.02 for cash tokens. And that price stays the same whatever time of day you use it. The M plan is Minimax's coding plan at $22.55 or $132 a month, and there's a discount down below. A llama cloud charges the same as the API, and Minimax code gives you free credits each day before it starts using your M plan. So, what is really the gotcha here? M3.1 runs at a maximum reasoning effort by default, which means it thinks as hard as it can before about everything, even simple jobs. That's why it sat there for up to 47 seconds before answering in the speed race, and why it cost about $0.01 a run there compared with about $0.01 for HighC. To prove it, I ran exactly the same inbox test again with the same 10 emails and the same 40 step limit, but this time I set the effort to low. It passed all 4 runs, and it finished about 28 seconds instead of nearly 2 minutes, which is a huge difference for 1 setting. The cost halved as well to about $0.01 a run, which puts it level with DeepSeek on time and roughly a 3rd of the price. My genuine Orange recommendation is to drop it to medium. Low passed everything in my inbox test, but from using it day to day over thinks on the higher settings. And medium still leaves it room for the harder jobs. So, to wrap up, Deep Seek won on the speed race and automation test, although M3.1 matched it on the automation once I turned the effort down. On video, M3.1 was the cheapest to get things right on what it could see, and Gemini is the one you'd use if you need sound. Across everything with a right or wrong answer, M3.1 was the cheapest per correct answer of the lot, and it tracked down and fixed a nasty crash in the app I built. MiniMax said this code's like the top models for a fraction of the price. I only tested it against other flash models, so I can't tell you it matches the very top, but for the price it more than held its own, as long as you turn down the effort. If you're building agents that need to read your screen, or you want to build that doesn't change with the time of day, I honestly think MiniMax have now got this one. If raw speed is all you care about, Deep Seek or Mercury will serve you better. If you want to use M3.1 in Claude Code every day, this is the video where I set it up in the coding plan and the discount code's down below. I'm so thankful for MiniMax sponsoring this video. Links to everything will be down below. Like, subscribe if you haven't already. Till next time, keep on vibing.