Claude Almost Bankrupt Me...
ฝัง
- เผยแพร่เมื่อ 7 ก.พ. 2025
- We all love Claude, but unfortunately it's 10-40x more expensive then the alternatives.
Check out T3 Chat: soydev.link/chat
SOURCES
x.com/theo/sta...
Check out my Twitch, Twitter, Discord more at t3.gg
S/O Ph4se0n3 for the awesome edit 🙏
Hey Theo! Small note: you may want to limit the number of messages users can make per chat. I work in this industry and almost all newer users don’t start a new chat when asking completely new questions, and you pay the price of sending their entire chat history as input tokens each time.
If you think of it like Teams or Slack or even texting, I think people imagine it like a person; they don’t start a new chat with their friend or coworker, they resume the previous chat with them. These users always costed us the most amount of money, so I thought I would mention it!
Absolutely right. Each question will send all message in the thread to backend relaying to LLM. it's kinda counterintuitive for end user who think there is a real conversation id to resume the memory.
In fact, this how Claude limits it in their web app. You'll see warning about daily limit if your convo is too long.
I think you can be really smart about it now though, maybe have gemini flash check if it’s an entirely new question or not before prompting claide for example
Or limit the lenght of history also
I always create a new chat with AI because I'm afraid that my new question may be negatively affected by the message history
will always love theo for being the youtuber who doesn't just talk stuff but also built it for commercial use. mad respect!
I've started watching theo around the time I found his Funfunfunction channel, roughly a month or two after he started releasing videos on it. The quality of the content and information provided is obvious for all to see. To quote a conspiracy theorist, funnily enough, "people can see the real humans from the snakes", or something like that, I don't remember the exact quote. To quote a saying from where I leave: Who knows, recognizes. (Those who know, recognize.)
I won't be using this product until it's renamed to "chat thing"
😂
Chat Thang
@@JamesRouzier Chat Thong
Fun fact: in websim, o3-mini-high actually costs a little less than Sonnet, despite superficially looking like it’s cheaper than Haiku 3.5. That’s because it generates almost 5x the output tokens! Keep an eye on reasoning model costs.
Also be wary of Deepseek R1 pricing. The official api is likely subsidized and about a quarter of our requests fail. Fireworks changed their pricing to $3/$8 input/output, but being a reasoning model will probably generate more output tokens and be a little less than Sonnet cost per message in practice
@@roberthaisfield3349 I'm pretty sure they're using Groq for Deepseek
I really appreciate your openness, Theo. This cost breakdown is incredibly helpful for me. I'm actually developing a similar macOS application, with a Swift Vapor backend. Based on your insights, I plan to implement a tiered pricing structure - one tier for lower-cost LLMs (under $1) and additional tiers for premium models like Claude. Your detailed explanation about this topic has helped me avoid potential customer dissatisfaction. Thank you so much! ❤
"today's sponsor is me talking about my thing and mentioning how I didn't use another sponsors thing for a sec" 😂
Nah Theo not being a shill impossible challenge
4:45 Gemini is hilariously fast and honestly it’s really good now
Yup - fully switched. Its ridiculous how good it is.
gemini really good right now
Pro?
@ yup 2.0 pro and the thinking models are lightning fast too
The fact that i can but a product that every change has a 15+ min TH-cam video telling why the change happened is so nice.
Im not in position of buying t3 sub (I dont have 8$ per month for ai) but if when ever i want such a product i will come here.
Maybe i can have some of my friends come and we buy a one account but fairly use it in our group
I really admire the level of transparency and openness you operate with. At the end of the day you're still running a business but the way you run it and communicate to your customers is so refreshing
My older dad is very interested in AI but most of the apps frustrate him on mobile. This he got used to. big W
T3 Chat really helped me to pick the best model for my AI-driven CAD for 3D printing. Before that I had to get like 5+ subscriptions to try all the premium models.
Just curious, how has a text-based model help with AI driven CAD? I haven’t looked into AI CAD yet, so forgive my ignorance
@@nebula391i am curious too
@@nebula391 I'm making an app that can do CAD using text/image input. Right now it is mostly useful for making custom organizers and has some presets for Gridfinity. So you can type something like "make a 1x1 gridfiinty bin for USB sticks", and it will look up USB dimensions and make a model with sockets and some reasonable space. It stores shapes in STEP and OBJ coordinates and outputs .stl files for printing. Right now, like 1/10 models is successful, so I just run each prompt 20 times and pick the best one. It is kinda hard to define shapes in text-format, but it can accept drawings and make models out of them. Can also do like Lego pieces and different gears, but the accuracy is even lower.
maybe openscad stuff?
Congratulations for the transparency and the project and everything! Really interesting stuff! Good work Theo / Mark and team!
I love the sponsor slot, and T3 Chat is great. Thank you!
Wasn't disappointed by a new policy at all. I hope your project succeeds in the long run. One suggestion - could you thing about the solution (not exactly in T3 chat), where users could request features and/or provide feedback? I know that your backlog is stuffed but would like to add an "anthropic's projects" feature to the pile! Thank you, appreciate your work
This research looks like consulting level quality. Well done!
This is a classic strategy to capitalize on low-volume subscription users. Companies attract customers with a low price, but when the reality of high-volume users sets in, they backtrack. It’s impossible to believe this couldn’t have been anticipated.
"Our competition kind of sucks" EZ
Just tried T3 chat. Holy crap it is responsive. Great work!
allowing us to buy more credits is the perfect solution. you gotta do what you gotta do
Ditto. A fixed subscription massively subsidies power users, and a credits system at least makes them pay for their usage.
Yep, windsurf (and I’m sure cursor) support this, when I have a very busy month and I’m doing a lot of prompting, I’ll often top up on credits. It’s a win win
When I saw the announcment that the rate limiting was going up. I kind of just laughed. I genuinely did not think what you were doing was sustainable, but I was happy you were trying to offer it.. even if unknowing of the potential damage
Hello Theo, I love the app it's just amazing ,I love the speed and also you motivated me to start exploring some ai models. Love from Romania
Riding the wave of techno-feudalism one token at a time
Theo got that "People like me, ensuring your packets get delivered, un-sniffed. So what do I do? I make sure that one bad config on one key component doesn't bankrupt the entire F company. That's what the F I do." Energy in this one
Thank you for building (and continuous improvement) a nice AI chat.
At that cost it’s almost worth hosting yourself for internal stuff.
Love the transparency, going to sign up now
NGL... been loving this Theo-dives-into-AI era. Thank you t3chat.
I've been a long user of SaaS AI through NovelAI. Aside from their self-hosted compute and no VC funding, their secret sauce is their frontend UI with per token probability. Your popularity is well warranted, but they too had to handle resource hogs since they're also on a plan system. By being so accessible, you've probably garnered more spammers than an a bare bones interface, which is a good problem to have, maybe?
You can make an abstracted roadmap though to help people understand the philosophy underpinning your direction.
Dude good stuff, glad to hear also you dont breeze through rate limit implementations, just learning how for a production app and its stressing me out
Update: Just got approved to move us to use upstash 😎
Add mobile! Been loving supporting this project, and anyone that doesn't understand scrappy startup business 101 for someone that cares about something, getting into a field that is new, doesn't deserve to use the app. Keep doing what you're doing.
i love these business lessons videos
I think a sub price as low as possible while still maintaining a reasonable profit margin and the option to buy extra credits for power users sounds really good. maybe a pinned thread with a quite small context window would be a good alternative for temp chats (since u keep history) while not exploding cost with millions of tokens. maybe even a slider that shows you how many words are being used for the request.
What you provide here is bigger than you realise. you're proving in real time that treating your users with respect, treating them like people, has itself got value.
Unfortunately if you price sub-$20, you will be used as arbitrage for Anthropic/OpenAI chat. Lot of startups learned that the hard way
I’m reverse engineering an api and tried o3 mini high on a tricky problem. 10+ messages later I gave up and asked Claude. It fixed it in one message. I tried Gemini pro on a long excel formula. 4 tries later, I give up and Claude one shots it. It’s worth it.
when i'm at the bottom of t3 chat, i shouldnt have to keep clicking the f'ing scroll to f'ing bottom button
Problem is, they're the only ones currently I at least kind of trust to not use my data for training, because they're the only ones clearly stating it's opt-in. Sure, the self-hosted ones are a good alternative now, but especially the projects and "fragments" features are really valuable to my use-case.
Isn't it bankrupted? Claude Almost Bankrupted Me...
You’re correct. Bankrupt without the “ed” is a little smoother, possibly a better title. Or Theo just spelt something wrong.
@@kyle8575 Yeah, it's a little misleading. It ends in a t, so it kind of sounds like a past-tense verb instead of a noun. But bankrupt is a noun.
So 100 people, who are watching this, have smirky smile on their face. Their thought is most likely, right now: "Upsiee" 😅
If you are spending this much i guess perplexity would be spending 10k dollars per day for claude 3.5 sonnet...i personally use up to 200 messages daily
can you make a video on the lack of entry level jobs available
He did th-cam.com/video/2PiM2e4RtTg/w-d-xo.html
@@CeruleanCrow oh wow thank you
And what would that do for anyone?
@@ark_knighthelp me cope
...so this T3 is like the Startpage of AI chatbots?
open webui + openrouter
Yeah
A video showing pros and cons of each of the models in T3 would be cool for us newbies
This was incredibly informative. I’m building something that requires quality models and will maybe get high usage. Think I’m gonna go with Gemini 2.0 instead of Sonnet.
Gemini 2.0 Flash is where it's at : top tier yet super affordable 🚀
As a free user I'm so glad there's another model to try now. Great! I hope it keep succeeding.
This is the pain of being early, difficult to manage this when there is so much volatility around pricing and new models becoming available.
This does not supprise me much. It can be very very hard to wrap third party services that charge based on consumption and your unable to directly one to one match that to your own customer billing.
If this was some kind of lock-in service where once you start using it there's a cost involved in switching to something else then I would be pissed about the price change, but seeing as I'm free to switch at any time (without any sunk costs to lose) of course I'm not going to get up in arms. I rubber duck with Claude all the time plus use it to get me started on learning new APIs, so I'll just use a Claude subscription seeing as I'll go over those limits. That's perfectly fine though, no skin off my back or yours. I'll miss your snappy responses and I'd still recommend your service to others. Be proud! 😄
Hey! quick tip: PLEASE do some in depth analysis of the cost of azure (not for a video, for your internal usage). if you get a sponsored subscription from Microsoft they wont show you the operating price, making you use resources that might cost more to run than neccecary. this will bite you in the tail when you are not sponsored anymore
can we get like feature when we highlight a part of the output and you get that option to ask t3 to explain the highlighted part, something like chatgpt has.
I think most people are reasonable and don't expect you to lose money.
Plus its great value and a great product.
I need some guidance from anyone tbh. From a frontend perspective what does it take to build something like this? I mean for formatting and all that do we use markdown? and how to make a pre/code block look like that? I couldn't find a proper guide for like making a good frontend of a chatbot or a platform similar to GPT. It would be great if someone can help or guide me towards something.
Not sure is Theo is using it here but Posthog has an LLM observability feature in Beta, using it currently for an agent we are looking at deploying. Includes a cost calculator and can relate it to a user, that plus alerts is a good way to monitor usage/improve your proompting
I hope Claude will be history in next month!!! Very good but very very pricey! So I hope we will never need them again very very soon!!!!
Only thing keeping me on Claude is the projects feature. I'm switching to t3 chat once y'all launch a similar feature!
Theo has already looked at his AWS Bedrock API, it's a third cheaper than Claude's official API and his speeds are very good, I bet you can cut that bill in half
I would probably switch over if you added projects/context storage like Claude has, also having artifacts with versions.
ok, 8$ vs 20$ on claude..what is the catch? just curious, ill pay today for it😊
Credit System is a good choice, maybe a model info page which model is good to use for ...
Why not just allow api key usage if you can't handle the pricing model?
Theo, t3 chat is a great product.
So this is what AI costs... huh, been relying only on my 4090, so this was quite interesting to see. ^^
I thought the DeepSeek was substantially cheaper, but it didn't seem like it from the table shown.
They recently raised the price
The only reason I use Claude is the Projects feature. Other platforms dont have that yet. If you implement that well, with all models compatible, you just win.
what does the "project feature" do? asking because i dont use claude.
@@jonasvq Allows you to upload files into a project knowledge base. Currently I upload my entire medium-sized Django project, and just prompt "Got this error stack: [error stack, which includes file names]", and it can fetch relevant context from the correct files and actually fix bugs and stuff. It's pretty next level. Not sure if Cursor etc can do that, but I'm not sure I want it in my IDE anyway.
@@yeetdeets Why not just use an AI system that runs locally in that case? For example, Jetbrains IDEs have local models that are trained on your code.
respect! but don't go bankrupt!
Charge for more!
Hi Theo! I'm a new fan. You randomly popped up on my feed, and I subscribed a few weeks ago.
I've always wondered: how do you make money? Your bio says you're a TypeScript developer, yet you manage to pump out videos every day. Don’t you need sleep?
I film them all 1 day per week (live on wednesday). Make money from sponsors and from the company I run. Have teams on both sides helping make it all happen :)
@t3dotgg thanks for the reply, really appreciate it :) awesome stuff, keep up the good work.
Did i just develop respect for theo?
have a "New chat" prominently displayed similar to how DeepSeek does it (but not as intrusive), currently you have it in a pretty inconvenient spot and it also disappears when sidebar is collapsed - because like the top comment here says, a lot of cost is due to users not starting new chats when topic changes. Chat messages must not be limited in any way, this is just annoying (meaning also a lot of context must be kept).
Theo, thanks for the video!
Why Claude 3.5 Sonnet and not Haiku? Haiku is the “mini” version from Claude, and it costs 10 times cheaper!
One Claude masochist used $500 of Claude credits in two days and basically complained that wasn't enough. Wow dude, did you really think you found the infinite Claude usage glitch? What an asshole.
They probably just weren't aware of the cost :D
Honestly stunned someone can use $500 in Claude in a single day.
What a power user using claude actually looks like. What kind of work are they doing to use it this much extensively.
Are you guys going to add things like Claude’s Projects to T3 chat? Or something similar
How products like Cursor manage to handle the cost of Claude? They are just fixing this with VC money? No way $20/mo can cover all those unlimited thicc composer requests there.
Please add folders and ways to organise past chats, then I would subscribe to t3 chat.
That's why I hate cloud in general. They can charge like anything. On premise can only increase electricity as unknown cost 😅
You have to fix the issue with disappearing prompts. Every now and then I hit new chat, type a prompt hit enter and nothing happens. I don't get the loading spinner in the ui and but my prompt GOES AWAY. This is very annoying and almost makes me cancel the other day.
I like the credits system tbh. It’s better for both company and users.
if you need to pull back because of thing you are providing is bankrupting you then I think it's resonable to do so
Is there any possibility that you could offer a 20 dollar membership for 3000-4000 chats, and a 200 Claude limit?
What did you think would happen? People pay the $8 and lose the logins?
Not a wise idea to use Claude API for general stuff; it's very much capable yet it's very expensive and the context window is quite small. I'd usually use chapter LLMs like DeepSeek, Gemini, even ChatGPT to do the work then make Claude the arbiter or verifier of their answer; that's much cheaper, but not in real time.
should get bulk discount from anthropic heard windsurf was doin that
I appreciate the transparency! Now show your profit 😜
Only thing keeping me on claude is the project context. Do you plan to implement that?
I think people have to pay for and supply their own api keys. I'd rather pay for what I want and not have to worry about the difference for me.
I don't get the premise- you are just a middleman? If this does get popular in 2 seconds the AI providers can completely destroy you by modifying their rates/stopping offering you their services etc etc. Essentially this is a middleman that competes on speed, which is only really feasible right now because you have a lot less traffic? I am perplexed by what the proposition is here?
you’d be surprised but these providers don’t get their profits from the chats. Their profits are the API usage. and, t3chat uses their API, which gives them more money, so there is no incentive for them to shoot everyone as well as t3 chat lmao
I think his main value proposition is UX and aggregation of the models.
@@EpicSquare I tried it out, and it's a fairly stock standard chat interface. I tried latest GPT recently and it was really cool how it jumped into different layouts for coding and stuff. Any innovation theo makes can be adopted by the OG provider, there's little to no defensibility here.
@@TheOnlyTima T3 will need to cover the cost of the API call, plus add some margin for profit and expenditures. The API provider also offers the service direct through their own premium tiers. If they wish to, they can make the API charge much higher and reduce their own prices to completely freeze out T3 from being able to operate and eat all that margin for themselves.
My understanding is that the main use case for them offering these API's is so that you can build some application on top that can use them to solve specific issues, like running customer service desks, or utilising the LLM to perform some background logic like dynamically interpreting user data etc. Simply repackaging the chat itself does not seem to add any value unless it is somehow cheaper/faster/better UX, and all these qualities can be completely assimilated by the original provider without much effort. My guess is Theo knows this and his exit strategy is to sell to one of the LLM providers. The only real point of difference they can't copy directly is model switching, but really if you use these models you will quickly pick a favourite, so long term model switching seems dubious to me after you have tested them all...
claude sonnet is unbelievable good
I'm amazed you still use Twitter given its current death spiral...
Are you planning to add Gemini 2 flash? It's pretty fast and cheap.
It would be nice if you offer usage-based pricing. I would be happy to pay you 10% to 20% over cost for offering me a single point of interaction with the models.
Basically becoming a model interaction platform that bills based on actual cost Plus.
Man when will we get an Anthropic updated model?
14:20 T3 Chat is NOT successful because the competition sucks lol it's "doing well" because you are actively losing money. Cheap pricing is the only thing that sets you apart from others.
Hey Theo, Do you have this AI Model pricing table somewhere online? It would make it a lot easier than going into each provider to check their costs.
well, you found out, lol. pioneering could be expensive
Just tried using T3 chat. Received errors.
First: Stream failed for unknown reason
Then second: Uncaught exception
Wait for me when I activate prompt caching, o3-mini costs the same as Sonnet
This is how businesses work some users make you money some users lose your money. This idiot doesn’t know that he expects to make money with every user.