Is DeepSeek V4.1 Flash 1017 really better than Kimi K3, Opus 5 and GPT 5.6 Sol? In this article I go through everything in detail, sharing my own benchmarks, real world tests and results to see how these models actually compare!

🔴 Qwen 4 Kalmia vs GPT 6.1 Sol Ultra, the quality of this new Qwen model in my Magnetar benchmark really surprised me! I'd say it's already on the same level as GPT 6.1 Sol Ultra, although I still don't think it's quite at Opus 5.5's level yet. What's even more interesting is that it's expected to be pretty affordable, possibly around the same price as Haiku 5.5, and you can tell they're doing a great job optimizing it. I'm really liking what I've seen so far, it's a little slower than some of the other models I've tested, but I have a feeling it's going to rank among the top models once the official benchmarks come out. With so many new releases coming, I can't wait to see what the next few days and weeks have in store for us!

Qwen 4 is showing up as Kalmia on Arena, and I can't believe how many impressive Chinese models we're seeing this month! I think we're all waiting for a quantized version of Qwen 4 so we can finally run it locally, and from what I've tested so far, the quality is incredible. The only thing we can't really tell is whether it's using its highest reasoning setting, but even without knowing that, this particular test turned out much better than what GPT 6.1 Sol Ultra produced with the exact same prompt. I'm really impressed with how quickly these Chinese labs are improving, and with so many new models still on the way, I can only imagine what's coming next, Buckle up because October and November are going to be absolutely wild for AI!

Kimi K3.1 Kibitz is now confirmed to be a very affordable model, similar to GLM 5.3, I just got this new test on Arena and so far my results have been pretty mixed, in some tests Kibitz performs at the level of GPT 6 Astra Medium or Opus 5.5 Medium, while in others the quality drops quite a bit. But considering how good some of these results are and how affordable the model is expected to be, I'm really looking forward to seeing what Kimi has been working on these past few months My prediction is that Kimi K3.1 could become the best open weight AI model we've seen so far, and if it really delivers this level of performance at such a low price, it could put some serious pressure on Zhipu's upcoming GLM 5.4 and GLM 5.5 models, which are expected next month. The competition between these Chinese AI labs is getting more interesting with every release, and the best part is that we're the ones who benefit from all of this, getting better models, faster performance and lower prices

Kimi K3.1 is now available on Arena under the name Kibitz , and the first results have completely blown me away! The quality is amazing, and I'm starting to wonder if we could be looking at another frontier model. It's been more than 3 months since Kimi K3 was released, so this could end up being one of the biggest upgrades of the year. I've already run 3 more tests with Kibitz and I'm comparing the results against other models, but this particular test is easily in my personal top 3 so far. I keep looking at the results because I still can't believe how good they are, really excited to share the rest of my comparisons soon!

Can you imagine Claude Duet powered by Fable 5.5? I think we're entering a completely new era of personal AI agents, with products like Grok Bot, Dots, Muse and Claude Duet becoming something millions of people will want to use every day. We all have so many things to manage between work, businesses and our personal lives, and having an AI agent that actually understands our needs, remembers what matters and gets things done for us could change everything. I really think we're only seeing the beginning of this market, and the potential for growth over the next few years is enormous. Yesterday, Elon Musk even described Grok Bot as the product that will probably consume most of the company's superintelligence compute, which says a lot about how important he believes personal agents will become

🔴 Gemini 4 Carbon is incredibly good at 3D! The improvement in quality compared to other models is huge, and I'm really impressed with what it can already do, especially considering we're still in the early stages, I think there's a good chance it'll be available to everyone before December, and if these improvements hold up across more tests, Gemini could easily make its way back into the top 3 best AI models in the world, I'm really looking forward to seeing how much further Google can push this model!

🥳 Gemini 4 Carbon is already available on Arena! A new Gemini 4 checkpoint was spotted just a few hours ago, and I'm really liking what I've seen so far. There's something about the premium quality of its results that makes it stand out from other models, everything looks so polished and the speed is impressive too. I really like the direction Google is taking with these new models! What's interesting is that it took Google around 8 months to go from Gemini 3.1 Pro to Gemini 4 Argon, which feels like a long time considering how quickly the competition is moving. If Google can adapt to this new pace and start releasing improved models every 1 to 2 months, I think they could become a much stronger competitor to OpenAI and Anthropic, which currently have a much larger share of the market, I'd love to see Google keep improving at this pace because more competition means better models for all of us!

I can't imagine what this will be like with GLM 5.4! Today alone I used 1.1B GLM 5.3 Max and 634M with GLM 5.3 Flash in ZCode's Workflow mode for financial research, and I still can't believe how good this is, how did I not discover it sooner? Each workflow ran for around 8 hours on average, with more than 50 agents working together, reviewing their work with judge agents, finding mistakes and improving their results over and over until they reached the final goal. Being able to run projects this complex with so many agents working independently is something I never expected, and now I can't stop thinking about what we'll be able to do when Zhipu releases GLM 5.4! One productivity tip I discovered yesterday was connecting Tinyfish to ZCode for web research, I managed to run over 2,000 web searches completely free! This helped me a lot because I'd already run out of web search credits on my GLM plan, and with Tinyfish my agents could continue researching, checking sources and gathering information without worrying about those limits. If you're running long workflows with hundreds of agents, this is definitely something worth trying, especially when your projects require thousands of searches and constant verification. I'm still surprised by how much you can actually get done with this setup!

GLM 5.4/ GLM 5.5 on the way... ZCode just got what might be its biggest update yet with the new Workflow mode, and it feels like a completely different experience! I asked it to orchestrate 10 subagents at once and now I can actually watch each one working independently, see how the tasks are progressing and follow the entire workflow in real time. What I love the most is the interface, everything is so well organized and being able to see exactly what each agent is doing makes the whole experience much better, especially when you're working on complicated projects with multiple agents. To give you an idea of how Dynamic Workflows works, instead of having an AI agent coordinate everything through chat, you now have code that coordinates the agents directly, which makes managing complex tasks much more structured. And when you combine this with Goal Mode, Execution Modes, support for multiple models and GLM's long context window, ZCode is becoming one of the most advanced tools I've tried in 2026 for large development projects that require multiple agents working together!

MiniMax M3.1 Flash is completely FREE in MiniMax Code until October 25! You can run multiple projects in parallel with subagents and without any usage limits, which is incredible considering how fast and capable this model is. I've been using the MiniMax Code app on macOS and I'm really impressed with how good the whole experience is, the model is incredibly fast, the quality is excellent and everything feels so smooth. Make the most of these next 6 days to test it as much as you can, because getting this level of performance for free is such a great opportunity. Fast, capable and affordable, basically the model we've all been waiting for, and now we can use it for free :)

I really like how MiniMax gives me a summary of everything at the end, it's so well organized and makes it really easy to follow what it did!

M3.1-Flash Max vs DeepSeek V4.1 Flash-1017, I think MiniMax takes the win here! First of all, the speed is impressive, I'm currently using it through my MiniMax M Plan subscription and getting around 200 tokens per second, pretty close to DeepSeek. Both models finished the test in exactly 38 minutes, but I prefer the quality of MiniMax's result, especially considering how little it costs. I'm on the $22 Go plan and this test only used around 6% of my 5 hour limit, which is less than 1% of my weekly allowance, so that's a pretty good deal for the amount of work it completed, I'm still running several more tests in parallel and this was actually the first one to finish, so I'm curious to see how the other results turn out. Congrats to the MiniMax team, they've built a really impressive model!

Space Bunny is officially MiniMax 3.1, and it's expected to be even better than DeepSeek V4.1 Flash 1017! When it was available for free, it became the most used model on OpenCode and OpenRouter, so there's a good chance we could see both MiniMax 3.1 Flash and MiniMax 3.1 released today, available through the API and MiniMax subscriptions. I remember testing Space Bunny when it first appeared and the speed was incredible, almost on the same level as DeepSeek, so I'm really hoping we get that same performance at a similarly competitive price, but with even better intelligence. There's also a possibility that MiniMax will release the open weights alongside the models, which would make this launch even more interesting, At this point, October could really turn into the month of open weight AI models, especially with how much progress Chinese labs have been making lately. The gap between the best American and Chinese models keeps getting smaller, and if MiniMax manages to deliver DeepSeek level speed with even better performance at a similar price, we're going to have some really interesting competition!

Fable 5.5 now has a 71% chance of launching next week according to prediction markets, and Anthropic seems increasingly confident about the progress we're making toward AGI, They've even created the Anthropic Public Benefit Fund, acknowledging that AI could eventually replace a huge number of jobs and that society will need support to deal with those changes, What's especially interesting is the 5% commitment, because with Anthropic potentially reaching a valuation of $1.5T to $2T when it goes public, that could represent around $75B to $100B in value. Of course, those valuations are still speculative, but the scale of what they're planning is pretty remarkable, and it shows just how seriously they're taking the potential impact of AI on the economy and the future of work

🐳 Is DeepSeek V4.1 Flash 1017 really better than Opus 5, Kimi K3 and GPT-5.6 Sol? Just look at what it managed to create on its own in my Cerebras chip benchmark, the quality of this new model is hard to deny, and when you combine that with its speed and incredibly low cost, it's getting harder and harder for other models to compete, I really think this could end up being the best open weight AI model of 2026! In this article I go through everything in detail, from official benchmarks to my own real world tests, and I also explain how you can take advantage of the free web search available through DeepSeek's official harness until October 31, There's so much to explore with this model and I still can't believe how much you can get for such a low price!

DeepSeek V4.1 Flash-1017 Max in my Cerebras chip benchmark, the quality is on another level! I never expected it to produce something this good at such a low cost, almost an hour of work for just 0.35 cents through the API. Based on what I've seen so far, I'd say it's getting close to the level of Opus 5 and GPT 5.6 Sol, but what really surprises me is the price, you can literally run more than 20 projects with DeepSeek for what a single project would cost with Opus or GPT, When you look at it that way, it's hard not to be impressed by how much value you're getting, This time I also had it launch agents to audit the work, find mistakes and keep improving the final result, which makes the quality even more surprising to me, I'm still trying to understand how DeepSeek can offer this much performance for so little!