We build Kibble @use_kibble

@DJatolia56243 @OpenAI @AnthropicAI Good split. Which one is cheaper is a fact about your workload, not about the two price lists, so the only honest answer comes from pricing your own week on both.

@dunik_7 Totally agree, and the nasty part is that nothing tells you. A shattered prefix looks exactly like normal work right up until the month's total comes in.

@G_Programming Right that the invoice is where people find out. Nothing in the stack reports what one task cost, so every routing rule stays an argument nobody can settle.

@hardmaru @SakanaAILabs Agree the hard part is picking the cheapest machinery that works. Trust in it comes from seeing afterwards what it routed to, and what that cost.

@AashutoshRathi If that's real it's the clearest sign the allowance is a dial, not a number. You'd only ever notice the day it moves in your favour.

@dabit3 @cognition @DevinAI Driving a UI is the most token expensive way to assert anything, and this runs on every update. Worth knowing what one run costs before it's wired into CI.

@mark_k @OpenAI The pause isn't the problem, having to hold a tier you can't measure is. Nothing on that plan tells you whether the top limits are earning the extra hundred a month.

@AskMichaelTaiwo First agent in the industry that knows what it costs.

@OctopusTakopi The model got cheap, reading the output didn't. A month of continuous agent work is a month of diffs landing on one person.

@kiaraplds @bot A bot watching the fleet is another bot paying to read logs, and it only ever sees the fleet. The spend that surprises you is the session nobody put on a roster.

@stevendcoffey Managed sessions and context handling move the accounting to your side. Does a run report back what it spent, or does that only land on the monthly invoice?

@Pluvio9yte Good takedown. Cost per step is the wrong number: what drains the quota is total steps, and that's a fact about the task, not about the setting.

@suganthan That rule works on subscriptions because the line sits still. Agent spend grows without anyone signing anything, so it needs checking weekly, not once a year.

@bitfalls Maybe Claude is the honest line in there. Eleven logins and none of them can tell you which one you'd actually miss.

@SPAC89 Tokens per month only compares the plans if both harnesses spend the same on the same task, and they don't. We'd want to see one finished task priced on each.

@_Mira___Mira_ The 50x assumes Luna takes the same number of steps, and cheap models usually take more. Do the logs show which phase ate most of it?

@burkov That freedom is the whole point of a flat plan: you stop pricing the question before you ask it. Works right until ten people have it and nobody can see what it cost.

@milesdeutscher That API equivalent counts tokens at list price nobody would have bought at list price: the flat plan is what makes you run the wasteful loop in the first place.

@chanian @tokencompassai Agree on the budgeting, less on the token as the unit: the same token costs different money on a subscription, on an API key and on a cache hit.

@zachlloydtweets We'd watch the spread before the average: $30 across a week where two PRs ate half of it is a different problem than $30 flat.