Token based pricing generally makes a lot of sense for companies like Zed, but it sure does suck for forecasting spend.
Usage pricing on something like aws is pretty easy to figure out. You know what you're going to use, so you just do some simple arithmetic and you've got a pretty accurate idea. Even with serverless it's pretty easy. Tokens are so much harder, especially when using it in a development setting. It's so hard to have any reasonable forecast about how a team will use it, and how many tokens will be consumed.
I'm starting to track my usage with a bit of a breakdown in the hope that I'll find a somewhat reliable trend.
I suspect this is going to be one of the next big areas in cloud FinOps.
I wonder if first-party offerings like Codex and Claude will follow suit. Most "agents" are utter nonsense, but they cooked with the CLI tools. It'd be a shame to let go of them.
This whole business model of trying to shave off or arbitrage a fraction of the money going to OpenAI and Anthropic just sucks. And it seems precarious. There's no honest way to resell tokens at a profit, and everyone knows it.
I'm personally looking forward to this change because I currently pay $20/month just to get edit prediction. I use Claude Code in my terminal for everything else. I do wish I could just pay for edit prediction at an even lower price, but I can understand why that's not an option.
I'm curious if they have plans to improve edit prediction though. It's honestly kind of garbage compared to Cursor, and I don't think I'm being hyperbolic by calling it garbage. Most of the time it's suggestions aren't helpful, but the 10-20% of the time it is helpful is worth the cost of the subscription for me.
Good change. I’m not a vibe coder, I use Zed Pro llm integration more like glorified stack overflow. I value Zed more for being an amazing editor for the code I actually write and understand.
I suspect I’m not alone on this. Zed is not the editor for hardcore agentic editing and that’s fine. I will probably save money on this transition while continuing to support this great editor for what it truly shines at: editing source code.
Zed and Warp were two promising Rust-based projects that I closely monitor. Currently, both projects are progressing towards becoming a generic AI Agentic code platform.
For those of us building agentic tools that require similar pricing, how does one implement it? OpenRouter seems good for the MVP, but I'm curious if there are alternatives down the line.
> Token-agnostic prompt structures obscure the cost and are rife with misaligned incentives
Saying that, token-based pricing has misaligned incentives as well: as the editor developer (charging a margin over the number of tokens) or AI provider, you benefit from more verbose input fed to the LLMs and of course more verbose output from the LLMs.
Not that I'm really surprised by the announcement though, it was somewhat obviously unsustainable
Making this prediction now, LLM pricing will eventually be priced in bytes.
Why: LLMs are increasingly becoming multimodal, so an image "token" or video "token" is not as simple as a text token. Also, it's difficult to compare across competitors because tokenization is different.
Eventually prices will just be in $/Mb of data processed. Just like bandwidth. I'm surprised this hasn't already happened.
Prediction: the only remaining providers of AI-assisted tools in a few years will be the LLM companies themselves (think claude code, codex, gemini, future xai/Alibaba/etc.), via CLIs + integrations such as ASP.
There is very little value that a company that has to support multiple different providers, such as Cursor, can offer on top of tailored agents (and "unlimited" subscription models) by LLM providers.
I completely get why this pricing is needed and it seems fair. There’s a major flaw in the announcement though.
I get that the pro plan has $5 of tokens and the pricing page says that a token is roughly 3-4 characters. However, it is not clear:
- Are tokens input characters, output characters, or both?
- What does a token cost? I get that the pricing page says it varies by model and is “ API list price +10%”, but nowhere does it say what these API list prices are. Am I meant to go to The OpenAI, Anthropic, and other websites to get that pricing information? Shouldn’t that be in a table on that page which each hosted model listed?
—
I’m only a very casual user of AI tools so maybe this is clear to people deep in this world, but it’s not clear to me just based on Zelda pricing page exactly how far $5 per month will get me.
seems fine - they're aligning their prices with their costs.
presumably everyone is just aiming or hoping for inference costs to go down so much that they can do a unlimited-with-tos like most home Internet access etc, because this intermediate phase of having to count your pennies to ask the matrix multiplier questions isn't going to be very enjoyable or stable or encourage good companies to succeed.
I was just thinking this morning about how I think Zed should rethink their subscription because its a bit pricey if they're going to let you just use Claude Code. I am in the process of trying out Claude and figured just going to them for the subscriptions makes more sense.
I think Zed had a lot of good concepts where they could make paid AI benefits optional longer term. I like that you can join your devs to look at different code files and discuss them. I might still pay for Zed's subscription in order to support them long term regardless.
I'm still upset so many hosted models dont just let you use your subscription on things like Zed or JetBrains AI, what's the point of a monthly subscription if I can only use your LLM in a browser?
Zed was supposed to be the answer to Atom / Sublime Text in my opinion, and I kinda do want to use it as my main driver, but it just isn’t there yet for me. It’s shameful because I like its aesthetics as a product more than the competition out there.
Just this other day I tried using it for something it sort of advertised itself as the superior thing, which was to load this giant text file I had instantly and let me work on it.
I then tried opening this 1GB text file to do a simple find/replace on it only to find macOS run out of system memory with Zed quickly using 20gb of memory for that search operation.
I then switched to vscode, which, granted opened it in a buffered sort of way and limited capability, but got the job done.
Maybe that was a me issue I don’t know, but aside from this one-off, it doesn’t have a good extensions support in the community for my needs yet. I hope it gets there!
40 comments
[ 4.6 ms ] story [ 77.8 ms ] threadUsage pricing on something like aws is pretty easy to figure out. You know what you're going to use, so you just do some simple arithmetic and you've got a pretty accurate idea. Even with serverless it's pretty easy. Tokens are so much harder, especially when using it in a development setting. It's so hard to have any reasonable forecast about how a team will use it, and how many tokens will be consumed.
I'm starting to track my usage with a bit of a breakdown in the hope that I'll find a somewhat reliable trend.
I suspect this is going to be one of the next big areas in cloud FinOps.
I'm curious if they have plans to improve edit prediction though. It's honestly kind of garbage compared to Cursor, and I don't think I'm being hyperbolic by calling it garbage. Most of the time it's suggestions aren't helpful, but the 10-20% of the time it is helpful is worth the cost of the subscription for me.
I suspect I’m not alone on this. Zed is not the editor for hardcore agentic editing and that’s fine. I will probably save money on this transition while continuing to support this great editor for what it truly shines at: editing source code.
Saying that, token-based pricing has misaligned incentives as well: as the editor developer (charging a margin over the number of tokens) or AI provider, you benefit from more verbose input fed to the LLMs and of course more verbose output from the LLMs.
Not that I'm really surprised by the announcement though, it was somewhat obviously unsustainable
https://news.ycombinator.com/item?id=45333425
Why: LLMs are increasingly becoming multimodal, so an image "token" or video "token" is not as simple as a text token. Also, it's difficult to compare across competitors because tokenization is different.
Eventually prices will just be in $/Mb of data processed. Just like bandwidth. I'm surprised this hasn't already happened.
There is very little value that a company that has to support multiple different providers, such as Cursor, can offer on top of tailored agents (and "unlimited" subscription models) by LLM providers.
I get that the pro plan has $5 of tokens and the pricing page says that a token is roughly 3-4 characters. However, it is not clear:
- Are tokens input characters, output characters, or both?
- What does a token cost? I get that the pricing page says it varies by model and is “ API list price +10%”, but nowhere does it say what these API list prices are. Am I meant to go to The OpenAI, Anthropic, and other websites to get that pricing information? Shouldn’t that be in a table on that page which each hosted model listed?
—
I’m only a very casual user of AI tools so maybe this is clear to people deep in this world, but it’s not clear to me just based on Zelda pricing page exactly how far $5 per month will get me.
presumably everyone is just aiming or hoping for inference costs to go down so much that they can do a unlimited-with-tos like most home Internet access etc, because this intermediate phase of having to count your pennies to ask the matrix multiplier questions isn't going to be very enjoyable or stable or encourage good companies to succeed.
I think Zed had a lot of good concepts where they could make paid AI benefits optional longer term. I like that you can join your devs to look at different code files and discuss them. I might still pay for Zed's subscription in order to support them long term regardless.
I'm still upset so many hosted models dont just let you use your subscription on things like Zed or JetBrains AI, what's the point of a monthly subscription if I can only use your LLM in a browser?
But as a mostly claude max + zed user happy to see my costs go down.
Just this other day I tried using it for something it sort of advertised itself as the superior thing, which was to load this giant text file I had instantly and let me work on it.
I then tried opening this 1GB text file to do a simple find/replace on it only to find macOS run out of system memory with Zed quickly using 20gb of memory for that search operation.
I then switched to vscode, which, granted opened it in a buffered sort of way and limited capability, but got the job done.
Maybe that was a me issue I don’t know, but aside from this one-off, it doesn’t have a good extensions support in the community for my needs yet. I hope it gets there!