NEW DeepSeek V4.1-Flash-1017 is WILD!

Julian Goldie (0:00)

So DeepSeek just upgraded its budget model. It's called V41 Flash 1017

And on four agent tests, this cheaper model scored higher than Claude Opus 5 So a big agent job on it costs about 12 cents. If you run it at night, their time, it's about 6 cents. And Flash is DeepSeek's cheap, fast model. It's the one you use when you don't want to pay for the best. Cheap models aren't supposed to be the top models. This one did on the tests that matter most if you use AI agents for work. So what is the update and why should you care? DeepSeek released it on October 17th. The name is the date. It says 10-17. That's why it's called 10-17. Of course, it isn't a brand new model. It's the same V41 Flash that came out in September. Same size, same design. DeepSeek just trained it more and every single score went up. The best part is you didn't have to do anything. If you use DeepSeek on the website, the app or through the API, you already have it. It's swapped over on its own. Now, why does this actually matter to you? Well, for example, we actually built this game with it, and it can build like really visual stuff. This is a flash model, and it's competing with frontier stuff, right? So if you use AI agents to get work done, two things decide if they're actually worth it.

And those two things are, you know, does the agent finish the job, and how much does each job cost to run? And this update moves both of those in your favorite ones. I'll walk you through every number in a second. Some of the wins, some of the losses, and one of them should change how you pick a model for each job. So first, a problem that fixes.

Agents work in loops. They read something, they think, they use a tool, they check the result. Then they go again, over and over until the job is done. Every step in that loop uses tokens. A token is a small chunk of text. Think of it like a word, roughly. So you pay for every token, the model reads and every token it writes. So one agent job can read hundreds of thousands of tokens on a top model that adds up fast. And that's been the trade for a while now. You could pick a smart model that costs a lot, or a cheap model that gets lost halfway through a long job. Cheap models are fine for quick stuff, but giving them a job with 20 steps and they'd wander off. That trade is what 1017 goes after. Strong on long, multi-step jobs, short answers. So it uses fewer tokens and a price that's a small slice at the top of the models. Now, DeepSeek actually says it's a Pareto improvement. That's a fancy word. It got better without giving anything up. Higher scores, faster, shorter answers, nothing traded away. And let me run you through the numbers. And quick note before I do, because almost every score here is a percent. If a model scores 77, that means it scored 77 out of 100 tasks. Simple. And let's start with the one that I actually care about. It's called Automation Bench. This test gives the model a business process and checks if it can chain tools and apps. Together, to finish it, start to win. 1017 scored 60.5, called Opus 5 scored 50.3, Chipsy 5.6, Sol scored 45.8. So the cheap model is 10 points ahead of Opus 5 here, and almost 16 ahead of GPT. That's its biggest lead on the whole chart.

If you automate workflows in your business, this is the road to watch. Next up is Apex Bench.

And this test, the kind of work an analyst or a specialist gets paid to do. You know, a real multi-step professional task. 1017 scored 72.1, Opus 5 scored 65.8.

That's first place by about 6 points. And DeepSeek says that even on its lowest effort settings, it still gets over 72% on the task. For about 15 to 20 cents a task, that's DeepSeek's claim. So hold it loosely, but if it holds up, that's pro level work for pocket change. Next, Agents Last Exam. So this one is brutal on purpose.

Long, hard, agentic tasks. Every model scores in the 20s or 30s. 1017 scored 35.4. Opus 5 got 28.6. On a test, this hard, a 7 point lead is big. Then there's coding, DeepSWE.

The model gets a real codebase and a real bug report. It has to fix a bug so the project's own tests pass.

1017 scored 77 Opus 5 scored 74 GPT 5.6 scored 73 Top of the table again. So that's four wins over Opus 5 All four are agent or coding tasks. Now the biggest jump of the whole update, Terminal Bench. And this drops the model into a command line, the black text screen, and makes it do jobs on its own. Install things, move files, fix broken stuff.