If I understand the change correctly, messages in long chats will “cost” more of the usage now. It would be great if there were an indictor what the last message actually costed.
I also support this, the usage limits lost all their transparency, maybe add a toggle to show usage on main drawer and or somewhere on screen.
I don't blame the new usage cost calculation for messages it is still very user friendly but the current 3 Star system is very confusing especially after some model were premium models and others were not.
maybe add seperate indicators one color indicator for relative cost based on context length wich starts at bright green in new chats no matter the model and and goes towards red wich shows the max context length and notifies you at certain points and one when selecting the model wich is better fit then the 3 stars.
Mark Florkowski
Team
Working on this, the difficulty is in accurate estimation
Maroon Quesadilla
why do you need to estimate anything, you subtract the usage from our limits, so why cant you show how much you subtracted?
Mark Florkowski
Team
The post I am replying to was regarding the $$$ indicators not the message cost.
Maroon Quesadilla
sorry my bad should have read it more carefully👍
Red Matchbox
This new cost model is a bummer. It should show how many tries you have, because some chats only eat 65% even with a Gemini flash medium. Only one query leaves you with only 35% remaining. When you jump to a new chat, you do one query, you’re done, you have to wait 3 hours for a refresh. It should vary according to the chat model you’re using, but it doesn’t feel that way. I feels like a punish for using the tool not for abusing the use
Mark Florkowski
Team
I took a look at your usage, and it seems you were sending some messages in super long threads, or those with many/large attachments. each message in these threads would cost between 5-30 cents, so obviously that is not something we can support sending hundreds of on an $8 subscription.
If you reach out to support@t3.chat, I am happy to reset your usage for the month so you can try again, but use less long threads. The limits are very reasonable if you can manage to start new threads when a topic changes.
Apricot Hill
Is there a way to add a “consolidate” button or similar with warning that thread length has gotten to the point where its becoming cost inefficient?
Mark Florkowski
Team
Definitely something we are looking into adding
Harlequin Bunny
For long threads, is the cost fixed, or does it take prompt caching into account?
Mark Florkowski
Team
It takes caching into account when available.
Amethyst Lemon
To clarify: just seeing the cost in some arbitrary usage unit after the fact in the statistics for nerds would help!
Indigo Peacock
I’d really like some indicator of when a message is going to cost more, reading here I’m only just now learning longer threads make longer cost. I’d be splitting off into new threads and stuff if possible.
I like the new system, but it’s extremely hard for me to preserve my usage as I really don’t understand what the potential cost of the message I will send is. It’s especially hard to judge how much the length of the thread/model/search/attachment will practically change the drain
Purple Kangaroo
Agreed fully. I don’t need “accurate”, but token estimation is easy to give a rough estimate - if everything else is going to be obscured like “number of dollar signs” and “percent remaining of a hidden amount”, then “number of words multiplied by 4/3” is a sane enough estimate to be useful and could easily turn into a quick formula of “this is an inexpensive model + and you have a short prompt + but you have a long history = this is as if you are sending an initial short request to a $$$$ model” or even “clicking send is probably going to use 10-15% of your current rate-limit window”.
like, really, just an estimate. This isn’t something I need to be able to add to a cost analysis spreadsheet, that’s what APIs are for. I just need a good feel of “is it worth it to me to keep the context right now?”
Sapphire Geyser
Is the cost per token fixed? If so would it make sense to include something like the total number of tokens currently in your “bucket" (overage + 4hr)?
Mark Florkowski
Team
Each model costs a different amount per token. There is an indicator of their relative cost in the model selector.
Sapphire Crab
I would love that feature. Just remember - when you change topic (completely, not just shift it slightly), start a new chat. That way you’re not only improving accuracy of the model (it has lot of stuff biased due to what is in the context window), you’re also not sending entire previous conversation again (meaning you lose a lot less tokens).
New model (refreshing every 4 hours) is both great and…sometimes annoying. Few times when using Kimi K2.5 (without thinking option and often with websearch turned off), after 3 hours of programming on one topic, I found myself being 15% from the limit. Something that I would use in less than 30 minutes. Of course waiting an hour isn’t that bad - especially since you can send one or two messages in the meantime, but it is stressful. I can’t even imagine using models that cost more. And that was on one topic (programming Manim to make an animation). I discovered that my problem was sending back too many images of what was wrong (often 3 max would suffice, but sometimes I sent 10). So this system is reasonable, it’s just annoying that I have to keep 2nd tab open with usage and check it before or after sending a message. And the cost of message would be great. It’s just that every model has it’s own tokenizer, so count would be different for every model. Introduce multi-modality (and sometimes you don’t know the cost of image until it is tokenized) and agentic use of websearch and the calculations are extremely tough to make.
Log in to comment and vote
Comments17
Maroon Canteen
I also support this, the usage limits lost all their transparency, maybe add a toggle to show usage on main drawer and or somewhere on screen.
I don't blame the new usage cost calculation for messages it is still very user friendly but the current 3 Star system is very confusing especially after some model were premium models and others were not.
maybe add seperate indicators one color indicator for relative cost based on context length wich starts at bright green in new chats no matter the model and and goes towards red wich shows the max context length and notifies you at certain points and one when selecting the model wich is better fit then the 3 stars.
Mark Florkowski
Working on this, the difficulty is in accurate estimation
Maroon Quesadilla
why do you need to estimate anything, you subtract the usage from our limits, so why cant you show how much you subtracted?
Mark Florkowski
The post I am replying to was regarding the $$$ indicators not the message cost.
Maroon Quesadilla
sorry my bad should have read it more carefully👍
Red Matchbox
This new cost model is a bummer. It should show how many tries you have, because some chats only eat 65% even with a Gemini flash medium. Only one query leaves you with only 35% remaining. When you jump to a new chat, you do one query, you’re done, you have to wait 3 hours for a refresh. It should vary according to the chat model you’re using, but it doesn’t feel that way. I feels like a punish for using the tool not for abusing the use
Mark Florkowski
I took a look at your usage, and it seems you were sending some messages in super long threads, or those with many/large attachments. each message in these threads would cost between 5-30 cents, so obviously that is not something we can support sending hundreds of on an $8 subscription.
If you reach out to support@t3.chat, I am happy to reset your usage for the month so you can try again, but use less long threads. The limits are very reasonable if you can manage to start new threads when a topic changes.
Apricot Hill
Is there a way to add a “consolidate” button or similar with warning that thread length has gotten to the point where its becoming cost inefficient?
Mark Florkowski
Definitely something we are looking into adding
Harlequin Bunny
For long threads, is the cost fixed, or does it take prompt caching into account?
Mark Florkowski
It takes caching into account when available.
Amethyst Lemon
To clarify: just seeing the cost in some arbitrary usage unit after the fact in the statistics for nerds would help!
Indigo Peacock
I’d really like some indicator of when a message is going to cost more, reading here I’m only just now learning longer threads make longer cost. I’d be splitting off into new threads and stuff if possible.
I like the new system, but it’s extremely hard for me to preserve my usage as I really don’t understand what the potential cost of the message I will send is. It’s especially hard to judge how much the length of the thread/model/search/attachment will practically change the drain
Purple Kangaroo
Agreed fully. I don’t need “accurate”, but token estimation is easy to give a rough estimate - if everything else is going to be obscured like “number of dollar signs” and “percent remaining of a hidden amount”, then “number of words multiplied by 4/3” is a sane enough estimate to be useful and could easily turn into a quick formula of “this is an inexpensive model + and you have a short prompt + but you have a long history = this is as if you are sending an initial short request to a $$$$ model” or even “clicking send is probably going to use 10-15% of your current rate-limit window”.
like, really, just an estimate. This isn’t something I need to be able to add to a cost analysis spreadsheet, that’s what APIs are for. I just need a good feel of “is it worth it to me to keep the context right now?”
Sapphire Geyser
Is the cost per token fixed? If so would it make sense to include something like the total number of tokens currently in your “bucket" (overage + 4hr)?
Mark Florkowski
Each model costs a different amount per token. There is an indicator of their relative cost in the model selector.
Sapphire Crab
I would love that feature. Just remember - when you change topic (completely, not just shift it slightly), start a new chat. That way you’re not only improving accuracy of the model (it has lot of stuff biased due to what is in the context window), you’re also not sending entire previous conversation again (meaning you lose a lot less tokens).
New model (refreshing every 4 hours) is both great and…sometimes annoying. Few times when using Kimi K2.5 (without thinking option and often with websearch turned off), after 3 hours of programming on one topic, I found myself being 15% from the limit. Something that I would use in less than 30 minutes. Of course waiting an hour isn’t that bad - especially since you can send one or two messages in the meantime, but it is stressful. I can’t even imagine using models that cost more. And that was on one topic (programming Manim to make an animation). I discovered that my problem was sending back too many images of what was wrong (often 3 max would suffice, but sometimes I sent 10). So this system is reasonable, it’s just annoying that I have to keep 2nd tab open with usage and check it before or after sending a message. And the cost of message would be great. It’s just that every model has it’s own tokenizer, so count would be different for every model. Introduce multi-modality (and sometimes you don’t know the cost of image until it is tokenized) and agentic use of websearch and the calculations are extremely tough to make.
Still if @Mark Florkowski can do it - great!