GLM-4.7-Flash: 42x Cheaper Than Claude, Actually Good at Coding!

There's a new open-source model called GLM4. 7 Flash that just dropped and it's actually pretty interesting. It's doing really well on coding benchmarks. It's MIT license and you can run it locally. Today I'm going through what it is, the benchmark numbers, and whether it's actually worth paying attention to. So, let's jump in. GLM 4. 7 Flash is from

Z. AI, AI, which is a Chinese AI company founded by some researchers from Shingua University back in 2019. They've been putting out the GLM series of models, and this is their latest release. The model family has three versions. The full GLM 4. 7, a quantized FP8 version, and then this flash variant. Flash is basically optimized for the balance between performance and being able to actually run it without needing a data center. It's 31 billion parameters which puts it in that midsize category. Z. ai is positioning this as the strongest model in the 30 billion class which is obviously a marketing claim but the benchmarks are legitimately good. The model is specifically tuned for coding agent task and reasoning. And yeah, it's completely open- source under MIT license. So you can do whatever you want with it commercially. You can grab it from hugging face right now. And the model card has all the deployment instructions if you want to run it yourself. So I wanted to test this model

out and see like how good is it in actual production. So this is a really quick demo. Obviously I haven't deployed this model locally but I'm using hugging chat to kind of deploy it. But you can see that I ask it to create a voxal art environment featuring a temple with a garden and everything. And it called bunch of tools. The first one being gr17 image turbo generate and it created this temple that we see here. So this image model that it used to generate this is pretty good. Like these are highquality production of those images, animals and a temple. So this is really good. But so then I asked it to actually code this in one single HTML file just so it's easy to deploy. And we can click on the preview button here to see what the temple looks like. Obviously I would set the expectations really clear. This is not competing against Claude Sonic 4. 5 or GPT 5. 2. So the expectation when I look at this is not going to be that it's going to be a beautiful looking vauil environment, but something that actually works and I can deploy without burning so much token cost. Right? So let's take a look at this. So we can see this is our garden. We have basic abilities to move around. We can drag to rotate, scroll to zoom up and down. It added some animals. We're kind of seeing two bunnies over here which is pretty good. And then we see these birds which are kind of like flying saucers in the air which is not bad at all. Like the vauer environment as you can see like it's not bad. Like the trees look good. They're uniform. Added some height. And here's our waterfall. So I'm not going to like criticize this model a lot. It's good for what it's supposed to do which is get up and running and produce something that you can use. And you can see the coding it's done is all the way over here. It's like good amount of lines of code and it's doing all the mathematics for that boxer environment. So, this is not bad at all and it was able to do this pretty quickly. All

right, let's talk about the actual benchmarks because this is where it gets interesting. On Amy 2025, which is a math competition benchmark, it scores 91. 6%. That's comparable to what you would see from GPT models. For reference, coin 330 billion 83B thinking gets 85%. So, GLM is ahead there. GPQA tests graduate level science knowledge and GLM 4. 7 flash hits 75. 2%. Quen 3 is at 73. 4%. GPTOSS is at 71. 5%. So that's pretty solid for a GLM 4. 7. On the live code bench version 6, which is a coding benchmark, it scores 64%. Which is decent, but Quen 3 gets it 66%. GPT OSS gets 61%. So, it's still competitive, but not the top model. But here's where it gets really interesting. On the software engineering bench verified, which tests whether the model can actually solve real GitHub issues in real code bases, GLM 4. 7 Flash scores 59. 2%, Quen 3 gets 22% and GBT gets 34%. That is a massive gap. It's not just a little better, it's almost triple what Quinn is getting. Then there's the TA 2 bench which tests agentic capabilities like tool use and multi-step reasoning. GLM scores 79. 5% compared to Quent 3's 49% and GPT's 47. 7%. Again, substantially better. Browse tests web browsing capabilities and GLM gets 42. 8%. The next closest one is 28. 3. And then there's humanity's last exam which is a really hard reasoning benchmark. GLM scores 14. 4%. 4% which doesn't sound that impressive until you see that Quinn gets 9. 8% and GPT gets 10. 9%. This benchmark is just brutal for all models. So what does this tell us? The model is strong across the board but is particularly good at coding tasks and agentic workflows. The software engineering bench score is genuinely impressive. That's a benchmark where most models struggle and GLM is doing almost 60% which makes it useful in that territory. Now, let's talk about pricing

because this is actually pretty interesting. Z. AI offers both paid API access and completely free versions. For the paid models, GLM 4. 7 flash X is dirt cheap. The input is only 7 cents per million tokens and cash input is 1 cent and output is 40. Compare that to the full GLM 4. 7 at 60 cents input and $2. 2 output. Flash is onetenth the price for input and about 1/5 for output. If you look at the other models in their lineup, GM 4. 7 Flash is competing with models like GLM 4. 6V Flash X at 4 cents input at 40 cents output and GLM 4. 5 Air at 20 cents input and $110 output. Flash sits right in the ultra cheap tier while maintaining better performance than most of them. But here's the really interesting part. There are three completely free models. GLM 4. 7 Flash, GLM 4. 6 V Flash, and GLM 4. 5 Flash. These are free for API access and there are no rate limits mentioned, just free. Obviously, Z. AI is doing this to get adoption and gather data. But if you're building something and want to prototype without burning money, you can just use the free tier. The cash input pricing is worth noting, too. Once you send context to the API, subsequent requests with the same context are massively cheaper. For GLM 4. 7 Flash, cash input is 1 cent versus 7 cents for new input. If you're doing agentic workflows where you're sending the same system prompt or codebase context repeatedly, this adds up. For open source alternatives, you're usually running them locally. So there's no per token cost, but you have hardware costs instead. So if you're comparing GLM 4. 7 flash API pricing to running Llama 3. 1 locally, you need to do the math on your GPU cost versus just using the API. For a lot of people, 7 cents per million tokens is cheap enough that it's not worth running your own infrastructure. The value proposition here is pretty clear. You get performance that's competitive with much more expensive models at a pricing that's an order of magnitude cheaper with a free tier if you just want to

test it out. So bottom line, GLAM 4. 7 Flash is a solid model with genuinely good benchmark numbers, particularly for coding. The software bench score of 59% is legitimately useful and the pricing is extremely competitive and having a free tier makes it just easy to try out. So, if you're doing coding work or building agents, this is worth testing. The thinking modes seem to help with complex tasks, and the API pricing makes it viable for production use without worrying too much about cost. If you want to try the model out for yourself, I'll put the links in the description. Let me know in the comments if you end up actually using it. And that's it for today. Make sure to subscribe to our channel. We do real tests, not just headlines. Make sure you're also subscribed to the world of AI. And don't forget to check out our newsletter for deeper breakdowns you won't see on YouTube. and I'm growing my Twitter following. So, make sure you follow me on Twitter as well.