Zed's Pricing Has Changed: LLM Usage Is Now Token-Based

I'm a big fan of Zed but tbf I'm just using Claude Code + Nvim nowadays. Zed's problem with their Claude integration is that it will never be as good as just using the latest from Claude Code.

Though from everything I've read online, Zed's edit prediction model is far, _far_ behind that of Cursor.

When you get Claude Code's $20 plan, you get "around 45 messages every 5 hours". I don't really know what that means. Does that mean I get 45 total conversations? Do minor followups count against a message just as much as a long initial prompt? Likewise, I don't know how many messages I'll use in a 5 hour period. However, I do understand when I start bumping up against limits. If I'm using it and start getting limited, I understand that pretty quickly - in the same way that I might understand a processor being slower and having to wait for things.

With tokens, I might blow through a month's worth of tokens in an afternoon. On one hand, it makes more sense to be flexible for users. If I don't use tokens for the first 10 days, they aren't lost. If I don't use Claude for the first 10 days, I don't get 2,160 message credits banked up. Likewise, if I know I'm going on vacation later, I can't use my Claude messages in advance. But it's just a lot easier for humans to understand bumping up against rate limits over a more finite period of time and get an intuition for what they need to budget for.

Edit: To be clear, I'm not talking about Zed. I'm talking about the companies make the models.

https://forstarters.substack.com/p/for-starters-59-on-credit...

It already is. There’s been a lot of talk and development around FinOps for AI and the challenges that come with that. For companies, forecasting token usage and AI costs is non-trivial for internal purposes. For external products, what’s the right unit economic? $/token, $/agentic execution, etc? The former is detached from customer value, the latter is hard to track and will have lots of variance.

With how variable output size can be (and input), it’s a tricky space to really get a grasp on at this point in time. It’ll become a solved problem, but right now, it’s the Wild West.

The hard work is the high level stuff like deciding on the scope of the project, how it should fit in to the project, what kind of extensibility the feature might need to be built with, what kind of other components can be extended to support it, (and more), and then reviewing all the work that was done.

Agree with the sentiment, but I do think there are edge cases.

e.g. I could see a place like openrouter getting away with a tiny fractional markup based on the value they provide in the form of having all providers in one place

Why does not justify charging a fraction of your spend on the LLM platform? This is pretty much how every service business operates.

I've been exclusively using Claude Sonnet 4 model in VSCode and so far I've used 90% of the premium quota at the end of the months. I can always use GPT4.1 or GPT5-mini for free if need be.

It's based on qwen 2.5 coder 7B and you should be able to run it locally quite easily since it would only require about 8 GB of VRAM for the 8-bit version. Not sure if Zed supports this though, I'm not a Zed user myself.

For autoregressive token-based multimodal models, image tokens are as straightforward as text tokens, and there is no reason video tokens wouldn’t also be. (If models also switch architecture and multimodal diffusion models, say, become more common, then, sure, a different pricing model more tied to actual compute cost drivers for that architecture are likely but... even that isn’t likely to be bytes.)

> Also, it's difficult to compare across competitors because tokenization is different.

That’s a reason for incumbents to prefer not to switch, though, not a reason for them to switch.

> Eventually prices will just be in $/Mb of data processed.

More likely they would be in floatint point operations expended processing them, but using tokens (which are the primary drivers for the current LLM architectures) will probably continue as long as the architecture itself is doninant.

A tool that is good for everyone is great for no one.

Also, I think we're seeing the limits on "value" of a chat interface already. Now they're all chasing developers since there's a real potential to improve productivity (or sadly cut-costs) there. But even that is proving difficult.

I am really tempted to buy ChatGPT Pro, and probably would have if I lived in a richer country (unfortunetley purchase power parity doesn't equalize for tech products). The problem with Windsurf (and presumably Cursor and others) is that you buy the IDE subscription and then still have to worry about usage costs. With Codex/Claude Code etc., yeah, it's expensive, but, as long as you're within the usage limits, which are hopefully reasonable for the most expensive prices, you don't have to worry about it. AND you get the web and phone apps with GPT 5 Pro, etc.

This is her another reason why CLI-based coding agents will win. Every editor out there trying to be the middle man between you and an AI provider is nuts.

I'm pretty sure that's only while it's in preview, just like they were giving away model access before that was formally launched. Get it while it's hot.

As a corporate purchaser, "bring your own key" is just about the only way we can allow our employees to stay close to the latest happenings in a rapidly moving corner of the industry.

We need to have a decent amount of trust in the model execution environment and we don't like having tons of variable-cost subscriptions. We have that trust in our corporate-managed OpenAI tenant and have good governance and budget controls there, so BYOK lets us have flexibility to put different frontends in front of our trusted execution environment for different use cases.

TextEdit may be worth looking into as well? Haven’t tested it for large files before.