I Replaced Opus With MiniMax M3 In Claude Code (It's INSANE)

Income stream surfers| 00:10:10|Jun 1, 2026

Chapters8

Introduces the MiniMax M3 as a fast, multimodal foundation model that handles text, images, and video with a 1M context window.

MiniMax M3 inside Claude Code delivers blazing-fast multi-modal performance with huge context, and the plan price surprise could redefine cost expectations.

Summary

Income stream surfers' host dives into the MiniMax M3 release, showing off its multi-modal capabilities (text, image, video) and a remarkable 1,000,000 token context window. He demonstrates using Claude code router to test the model inside Claude Code, noting the performance uplift and substantially lower prefill/decode costs versus older MiniMax generations. While he’s skeptical about past MiniMax results, the M3 impresses with speed and design quality, even as he highlights a pricing shift: the normal price is reportedly higher than before, but a 50% off promo for the first seven days of June drops the cost to a referenced 0.30 with 120 output. He compares MiniMax to Opus and other Chinese models like DeepSeek Pro and Mimo, acknowledging token-efficiency and broad usage in ecosystems like Hermes. The creator openly discusses plan mode testing, cache quirks, and timing (early morning builds) while teasing continued testing tomorrow. He also plugs Harbor SEO.ai as his own tool and encourages viewers to test MiniMax themselves, calling the design “premium-level” and potentially the new best-in-slot model. The video blends real-time experimentation with market chatter, including references to Open Weights, Gemini 3 Flash, and Frontiers in model deployment, all while maintaining a candid, first-person narrative. In the end, he signals optimism that MiniMax could outperform Opus 4.x on design and cost, but leaves room for more validation. The piece closes with a tease of future tests and a sign-off that keeps the curiosity high for viewers.

Key Takeaways

MiniMax M3 is a multi-modal foundation model that accepts text, image, and video inputs with a 1,000,000 token context window.
Compared to earlier MiniMax generations, M3 promises roughly 1/20th the cost at 1M tokens with faster prefill and decode while maintaining quality on most tasks.
Claude Code Router is used to access MiniMax M3 inside Claude Code, illustrating practical integration for developers.
During a 50% promotional period in early June, the plan price is shown as 0.30 with 120 output, though the normal price cited elsewhere is higher (60 before the promo).
The creator compares MiniMax favorably against Opus 4.x and popular models like DeepSeek Pro and Mimo, noting token efficiency and design quality.
Early tests in plan mode show impressive early results, but cache-related issues (Next.js cache errors) may skew initial impressions until retests are done.
Harbor SEO.ai is promoted as a tool the creator built for SEO content, highlighting monetization and helping viewers assess value beyond the model itself.

Who Is This For?

Essential viewing for developers exploring multi-modal AI models, especially those evaluating MiniMax M3 versus Opus or Open Source options, or those curious about Claude Code integration and pricing promos.

Notable Quotes

"There is a brand new model. It has different modalities now. MiniMax just dropped. The MiniMax M3 is a multi-modal foundation model for MiniMax."

—Introduction to the M3 and its multimodal capability.

"A 1 million context window ensures for long horizon agency."

—Highlighting the large context window of M3.

"Just so you guys know, I'm using this inside Claude code, right? So, this is the one that I use."

—Describing practical integration with Claude Code Router.

"This plan has now finished. Normally, I don't do plans, but I was just curious what would happen if I did a plan."

—Commentary on testing the plan mode.

"This is insane. This is the best I've ever seen. That's like I mean this is pretty pretty well written."

—Expressing astonishment at the generated output and perceived quality.

Questions This Video Answers

How does MiniMax M3 achieve a 1,000,000 token context window and why does it matter for long-horizon tasks?
Can Claude Code Router efficiently integrate MiniMax M3 for text, image, and video inputs?
Is MiniMax M3 cheaper in practice than Opus 4.x, and what should buyers expect during promo periods?
What are the trade-offs between MiniMax and popular alternatives like DeepSeek Pro or Mimo in Hermes?
What should I test first when evaluating a multimodal model like MiniMax M3 in a real project?

MiniMax M3Claude Code RouterMulti-modal modelsOpenWeight/Open source AIOpus vs MiniMaxToken efficiencyHermes ecosystemDeepSeek ProMimoHarbor SEO.ai

Full Transcript

Okay, look at this guys. There is a brand new model. It has different modalities now. MiniMax just dropped. The MiniMax M3 is a multi-modal foundation model for MiniMax. It supports text, image, and video inputs for text output. A 1 million context window ensures for long horizon agency. If I have to read this phrase one more time, I don't know. I'm going to quit YouTube, I swear. Coding and tool use. It is built on MiniMax sparse tension MSA, which replaces full attention with KV. I don't know what any of this means. Um roughly 1/20th of the cost of the previous generation at 1 mil tokens with substantially faster prefill and decode while retaining quality across most tasks. Now, MiniMax for me has not really been that good um in in the past, but what I will say is this one does seem extremely fast. Okay, so yeah, like I said, this is a very very fast model. Just so you guys know, I'm using this um inside Claude code, right? So, you can do this by using Claude code router. So, this is the one that I use. There's obviously different ways to do everything, guys. I don't know why people feel the need to comment saying you can do this without Claude code router. Like, I get it. I just use Claude code router, right? So, just Google Claude code router if you do want to use Claude code with different models, including these Chinese models. Now, I thought I'd do something a little bit different. I'm just doing plan mode um just because I wanted to see what it came out with. Now, one thing just quickly to mention is there is currently uh 50% off the MiniMax provider for the first 7 days of June. So, that comes out at 0.30 and 120 output. It's already been used 2.2 billion It's already used 2.2 billion tokens, which is pretty crazy. And yeah, a lot of people using it inside Hermes. This is pretty common. Also, Pi, which is becoming more and more popular. I might have to test out Pi. But, I haven't had time just yet. Just while this is loading, guys, I might well just shout myself out. This is Harbor SEO.ai. This is my tool that I created to help people basically write optimized SEO content for their businesses for a very, very good price. It's 29 euros a month and you can generate a crap ton of content with that. These are the current stats. So, we've got 856 pages published, 500,000 impressions, 4,000 clicks, which is a very, very good return on investment for basically anyone. So, yeah, go and check out Harbor, guys. There'll be a link in the description and in the pinned comment. Okay. So, this plan has now finished. Normally, I don't do plans, but I was just curious what would happen if I did a plan. Uh so, the plan is now approved. It's going to start smashing this out. This might take a little bit of time. It is 5:30 in the morning. Don't ask. Um but yeah, I'll I'll probably wait for this to build and we'll see what happens. Okay. So, we're starting with about $15 worth of credits, guys. We'll see how much this actually costs by the end of the build as well. But yeah, guys, I have to say, I've been pretty impressed with the models from China recently. DeepSeek Pro, a lot of people didn't like it. I thought it was a really, really good model, to be honest with you. You can see people agree. Not only is this like ludicrously cheap, um it's actually a pretty good model. Another one is Mimo as well. I've been using Mimo a fair amount myself and I have to say another really, really good model. People agree as well. That's a lot of weekly tokens. Like, if you compare this to Opus 4.7, uh that's 2.38 trillion tokens, which I'm surprised by, honestly. I would not have expected that. I wonder I'll be Hermes, like, um controlling agent or whatever they're called. I'll blow code. Hermes, yeah. So, I I guess this is what people are using Opus 4, which is why there's so many tokens. So, yeah, looks like uh DeepSeek V4 Flash is the most popular. HYP3 Preview, that's interesting to see here. So, I probably have to test this out honestly. That is that is a lot of tokens. But yeah, you can see Mimo here Deep C V 4. These are kind of the models that most people use. Gemini 3 flash I am not surprised to see here at all. I would have expected to see GPT 5 Nano here potentially. Um but apparently not. This is definitely interesting. Yeah, step 3 7 flash. This one was just released as well. I did see this yesterday. Haven't had time to test this out but people like this model for Hermes again uh interesting. But yeah, just jumping back to MiniMax. This is what we are currently testing. I have not been massively impressed by MiniMax. This is actually MiniMax M3. Which do they normally put in there? MiniMax M2.7 Yeah, they do. But yeah, they've slashed the price. This is this will be the eventual price is $2.40. So they have increased the price which most people from what I understood their biggest thing with MiniMax was that it was cheap. So it looks like they have increased the price. So this is with 50% off. So the normal price is 60. So they've tripled the price. Uh and also double the output price. I'm not sure people are going to like that. I have to say it's used completely nothing so far in terms of money. Um so yeah, it's definitely very very cheap still. I don't know what this has come from. I don't know why this is here Nano banana. Seems a bit random. That might be Harbor actually. So yeah, another thing that people talk a lot about is the API and token plan on MiniMax. I'm not going to talk about it in this video but um apparently it's pretty good um value for money. And it looks like Opes 4.7 being beaten by Browse Comp. I don't know what that means guys. Honestly, I don't know how important benchmarks are. They're very often uh benchmarks for sure. And it does look like it's open weight as well. So the first open weight model with three frontier capabilities. Um I don't know if they've released it just now. Let's check. So, Mini Max M3 hugging face. Let's see. Yeah, I'm not sure. I don't know if this is it, but um Yeah, I'm not really a big open weights guy, to be honest with you. I will eventually get a Mac that can run them, but for now, I'm just kind of guessing. This looks like it might be it, but Yeah, they haven't posted it to the top of their their organization card just yet, so Let's see what Reddit has to say. So, another model that seems that has vision, seems cheap and efficient. Nice. The new step 3.7 flash has vision. Interesting. But, I don't see the weights or even a mention of the parameters. Anyone know more than me? Weights will be released in the next few days. Yeah, so I mean, people kind of agree with me with Mini Max. So, this is either way bigger than 250 billion parameters. Benchmarks to previously unachievable levels of success. A breakthrough that will be a landmark moment in the open weight space forever. It's definitely not that one. And then this guy says, "Every Mini Max model has claims to be sota among open source and has never gone anywhere." I do agree with this. Uh I haven't I haven't found like massive success with Mini Max. But, we will see. This seems to be doing a fairly good job. Um But, yeah, we'll we'll see in a sec how it how it's built. It's been running for 17 minutes. And yeah, the cost is uh very good. Very very good. It seems to be quite token efficient. That is what people generally do say about this model. All these models, Mini Max. Seems to be very very token efficient. Oh. What? That This is insane. This is the The This is the best I've ever seen. That's like I mean this is pretty pretty well written. Okay, wow. Yeah, um they might be cooking. They might be cooking. Now, I don't know if this is because I use plan mode. I might have to redo this test tomorrow when it's not almost 6:00 in the morning. Uh this is just um a cache error. Um I might have to retest it without the without plan mode, but this is looking pretty insane, I have to say. This is because it's still coding, by the way. It hasn't actually finished. Uh this is probably the best design I've ever seen. Um really premium-level design. The technical build is good, too. Um it just looks like it's not because it's it's getting hit with Next.js uh cache errors. Okay, so the only issue is if I press uh Italian when I'm not on the home page, it doesn't work, but that's actually a very very common error. Uh again, this is just cache errors cuz it's still coding. Guys, I can't believe how good this is. I need to test this out more because this is completely absurd how good this is. But yeah, for now I think I'm going to leave the video there. It's it's late. Uh this is worth testing, guys. This costs absolutely nothing. Let's just go to credits. Yeah, it cost less than 60 cents. Um and yeah, the the code quality here is amazing. One of the best I've ever seen design-wise, for sure. Guys, go and check this out for yourselves. MiniMax is cooking something over there, for sure. There's a few issues, you can see. Um but just yeah, overall insanely well done. Wow. But yeah, I mean definitely worth a check. Definitely worth a try. That is some next level right there. Really, really good quality. That's really nice. I'm going to do a few more tests on this tomorrow guys, but this might be the new best in slot model. And I would have to say the design here is better than Opus. The design here is better than Opus 4.8. That's not an exaggeration. Thank you so much for watching. Go and check out harbor.ai. If you are watching all the way to the end of the video, you're an absolute legend. I'll see you very, very soon with some more content. I need to go to sleep. Peace out.