Pinned Tweet
Your AI feature can be popular and still eat your margin.
I'm Di, co-founder of Omev.AI. I break down input, output, and cached token prices, cost per customer, and when cheaper models make sense.
Follow if you're building B2B SaaS that has to make money after the API bill.
ALT Di Reshtei in a dark textured blazer and black knit shirt, seated in a light-filled studio and looking at the camera.
Di Reshtei retweeted
your ai saas has a '$300 cac, $100/mo, 3-month payback' slide.
example: $60/mo goes to ai + other delivery costs. $40 gross profit means 7.5 months to recover $300, assuming the customer stays.
that 3-month payback was a bedtime story.
Di Reshtei retweeted
your 'free ai trial' has a paying customer: you.
say each trial burns $1 in ai costs and 5% convert. that's $20 in trial ai spend per new paying customer, before ads.
bragging about cheap signups while ignoring that bill is expensive comedy.
Di Reshtei retweeted
your saas a/b test can torch your cache savings.
sonnet 5.5 per 1m tokens:
5-minute cache write: $2.50
cache hit: $0.10
25x apart.
if each prompt variant goes cold before reuse, you're budgeting for hits and buying writes.
has a prompt test ever spiked your ai bill?
Di Reshtei retweeted
your $0.01 ai report can come with a $1 babysitting bill.
example: 2 minutes of human review at $30/hour adds $1 per report. 100x the model call, before other delivery costs.
what costs more in your saas: inference or human review?
Di Reshtei retweeted
your ai saas charges overages. great. did the customer pay?
the model bill is real even when their card fails.
track ai spend against collected revenue. cap unpaid usage.
otherwise you're financing someone else's token binge.
your $0.01 ai report can come with a $1 babysitting bill.
example: 2 minutes of human review at $30/hour adds $1 per report. 100x the model call, before other delivery costs.
what costs more in your saas: inference or human review?
Di Reshtei retweeted
Just moved to Chi葯in膬u 馃嚥馃嚛
looking for saas founders, indie hackers and ai builders nearby.
would love to grab a coffee, swap stories 馃.
dms are open.
Lets connect 馃槂
who鈥檚 building here?馃檪
Di Reshtei retweeted
claude's api can store your chat. your token bill didn't get amnesia.
message threads send less data, but claude still reads the full conversation. token pricing is unchanged.
how do you stop long customer chats from eating your margin?
your saas a/b test can torch your cache savings.
sonnet 5.5 per 1m tokens:
5-minute cache write: $2.50
cache hit: $0.10
25x apart.
if each prompt variant goes cold before reuse, you're budgeting for hits and buying writes.
has a prompt test ever spiked your ai bill?
your ai saas charges overages. great. did the customer pay?
the model bill is real even when their card fails.
track ai spend against collected revenue. cap unpaid usage.
otherwise you're financing someone else's token binge.
Just moved to Chi葯in膬u 馃嚥馃嚛
looking for saas founders, indie hackers and ai builders nearby.
would love to grab a coffee, swap stories 馃.
dms are open.
Lets connect 馃槂
who鈥檚 building here?馃檪
Di Reshtei retweeted
your ai saas has 'great margins' because startup credits pay the token bill.
cute. the credits run out.
price every plan with unsubsidized ai + other delivery costs. fix the math before the free money ends.
what happens to your gross margin when the credits hit zero?
Di Reshtei retweeted
your ai saas saves pennies on input, then writes a bloody novel.
sonnet 5.5: $2 per 1m uncached input tokens. $10 per 1m output tokens.
if customers need 3 lines, why pay for 30?
measure cost per accepted answer. test shorter output before changing your price.
claude's api can store your chat. your token bill didn't get amnesia.
message threads send less data, but claude still reads the full conversation. token pricing is unchanged.
how do you stop long customer chats from eating your margin?
your ai saas has 'great margins' because startup credits pay the token bill.
cute. the credits run out.
price every plan with unsubsidized ai + other delivery costs. fix the math before the free money ends.
what happens to your gross margin when the credits hit zero?
Di Reshtei retweeted
self-hosted ai is 'cheaper' until you price the empty hours.
example: a $1/hour gpu left on for 30 days costs $720. only 20 busy hours? $36 per busy hour, before ops.
compare with your api token bill on the same workload. the gpu charges while your users sleep.
Di Reshtei retweeted
your saas pays real-time token rates for a report nobody reads until tomorrow.
claude's batch api lists 50% off on supported models, with a 24-hour processing window.
that's an expensive way to feel productive. move flexible work out of the rush lane.
your ai saas saves pennies on input, then writes a bloody novel.
sonnet 5.5: $2 per 1m uncached input tokens. $10 per 1m output tokens.
if customers need 3 lines, why pay for 30?
measure cost per accepted answer. test shorter output before changing your price.