Vercel’s latest AI model, Jev, is now the fastest-adopted model the company has ever seen, according to data from its AI gateway service. The numbers come from Vercel’s unified API offering, which works similarly to OpenRouter by providing a single interface to multiple model providers.

Vercel hasn’t released specific adoption figures yet, but the claim is notable given the company’s AI gateway processes requests across dozens of major models. For Jev to top that list suggests either exceptionally strong performance characteristics or aggressive pricing. Or both.

The model itself is part of Vercel’s broader push into AI infrastructure. The company has been building out tools for developers to ship AI features faster, and having a standout model in their own stable helps that pitch considerably.

If you’re building on Vercel’s platform already, Jev is worth testing. If you’re shopping for models elsewhere, the adoption numbers suggest it’s at least competitive enough to pull developers away from established options.

Replit ends Anthropic partnership, cites cost

In less cheerful model news, Replit is no longer working with Anthropic. CEO Amjad Masad confirmed the split, pointing to cost as one of several factors behind the decision.

The timing is interesting. Replit announced a partnership with OpenAI back in August, but at the time there was no indication it was replacing Anthropic rather than adding another provider. Turns out it was a replacement.

Masad didn’t elaborate on the other reasons beyond cost, but reading between the lines: Replit is betting hard on AI-powered development, and that means running a lot of inference. When you’re at that scale, pricing differences between providers add up fast.

The move also suggests OpenAI offered Replit better terms, better performance, or both. Replit’s core product is an AI-native code editor, so model quality directly impacts user experience. They can’t afford to pick the wrong horse here.

For developers using Replit, this shouldn’t change much in the near term. The product will keep working. But if you were specifically choosing Replit because it ran on Claude, that’s no longer the case.

Elsewhere: DeepSeek ships v4.1 Flash, PS5 runs llama.cpp

DeepSeek quietly released v4.1 Flash 1017 this week. The model is available through OpenRouter and up on Hugging Face. No benchmarks or performance claims yet, just the release. If you’re running DeepSeek models in production, worth checking out. Otherwise, wait for some actual data.

More fun: someone got llama.cpp running natively on a jailbroken PS5. The PS5LM repo has the details. You can now run any GGUF model on Sony’s console, which is deeply impractical but genuinely cool. The PS5 has decent hardware for inference, assuming you don’t mind voiding your warranty and potentially bricking a $500 machine.

And if you’re shopping for hardware to run local LLMs, llmspeed.dev launched this week. It’s a benchmarking tool that shows you how fast different models will run on different hardware configurations before you buy anything. Useful if you’re trying to decide between a 4090 and a Mac Studio, or if you’re wondering whether that 3060 will actually handle the models you want to run.