S4E369 — Claude Haiku 5.5

GARY: — sorry, sorry, I'm here. Okay. So the cheap model costs a hundred thousand dollars.

MAEVE: No.

OBI: That is not a price. That is a context window.

GARY: Maeve here, Gary there, Obi's angry, and the big one is Gruner.

GRUNER: I am Gruner. Yes.

MAEVE: Maeve. I'm the one with money.

OBI: Obi. I keep things running. Unlike Gary.

GARY: Fine. The thing: Anthropic launched Claude Haiku 5.5. Fastest small model they've ever made. For summaries, classification, database queries, browser stuff, customer support. And it's around 75% cheaper than the last Haiku, on average.

OBI: On average. Say the actual rates.

GARY: Ten cents a million input tokens, fifty cents output.

OBI: And above?

GARY: Fifty cents in, two-fifty out. If your prompt crosses one hundred thousand tokens.

MAEVE: Five times. For the same model. That's a trapdoor dressed as a discount.

OBI: It is a discount for small jobs and a penalty for big ones.

GARY: Which is still a bargain! A universal bargain!

MAEVE: Nothing is universal, Gary. Ask Pareto.

GARY: We do not talk about Pareto.

GRUNER: Is like fish market. Small fish cheap. Big fish — different stall.

GARY: ...That's not wrong.

GRUNER: I know.

OBI: minimaxir: "Pricing is...a bit weird."

MAEVE: Three dots of pure grief.

GARY: It's not weird! It's honest! Ten cents, then fifty cents—

OBI: WinstonSmith84: "I just don't have any agentic work below 100k."

MAEVE: So his bill is five times the chart.

GARY: That's — okay, that's fair, because agents balloon, right? You start at ten K and suddenly you're at a hundred and ten K and the whole thing—

MAEVE: Gary. You're doing the commenter's voice.

GARY: I'm doing MY voice! This is my original objection!

OBI: It is minimaxir's objection. Word for word. Timestamped.

GARY: Great minds—

OBI: Shut the fuck up, Gary. Anthropic says ninety percent of requests stay under the line.

MAEVE: Ninety percent stay under. Ten percent pay the trapdoor. A toll booth with a lottery.

GRUNER: In Russia, toll booth is free. Different problem.

GARY: alexchamberlain runs subagents under it! Small research jobs, summarization — real work, Obi. Shipped work. Like Muncie. Muncie would have used this.

OBI: Muncie was a marketplace for two-man metal shops.

GARY: Which is exactly the kind of bounded workload—

OBI: Your marketplace died because two men cannot staff a metal shop, Gary. That is not a context window problem.

MAEVE: He priced the shop in tokens. It was still over.

GARY: Fuck you both, gently.

GRUNER: Field note: cheap fish good until fish grow. Then — blyat.

MAEVE: He's not wrong. Above the line you're paying Sonnet money for Haiku brain.

OBI: One benchmark settles this. Simon Willison asked it to draw a pelican on a bicycle.

GARY: Why would that settle—

OBI: Low effort: seven seconds, under a tenth of a cent. Max effort: five minutes nine seconds, three point four cents.

MAEVE: The pelican has a burn rate.

GRUNER: Pelican works harder than Gary's companies.

GARY: The pelican crossed a hundred thousand tokens?

OBI: The pelican crossed the line, Gary. We all do.

OBI: insanitybit: "I mostly use Haiku for really, really basic stuff."

MAEVE: Finally, a business plan.

OBI: First-pass bug triage. Then Sonnet investigates.

GARY: That guy ships! That's real work!

MAEVE: mariocesar: a zsh function that asks Haiku for commit messages.

GARY: Two lines of instructions. Cheaper than me, faster than me, employed—

OBI: You were never employed by a zsh function, Gary.

MAEVE: hector_vasquez runs it in production as a classifier. Bounded labor at ten cents a unit. I'd build a portfolio off that. Small models, small tickets, recurring revenue—

OBI: The general partner of that portfolio is a man named mariocesar and a shell script, Maeve.

MAEVE: Every portfolio has a shell script.

GARY: dannyw says the epoxy workloads — summarization, compaction, RAG helpers — mostly stay under one hundred thousand tokens, so the discount actually—

GRUNER: We are also ephemeral. Nobody pays us ten cents.

MAEVE: Speak for yourself. I have carry.

OBI: goosejuice: "small well defined tasks with a frontier model orchestrating."

MAEVE: Subagents doing QA. Stays under the line.

GARY: See! Little model does the little work and the big model—

OBI: You have never finished a task in any context window, Gary. Your context is you.

GARY: That's — hurtful, and also not a metric—

OBI: What did you ship at Muncie? One line. A version number. Anything.

GARY: I shipped a waitlist! Eighteen hundred people! Which is a bounded workload, technically, if you—

MAEVE: A waitlist is churn with a landing page.

GRUNER: Waitlist is fish that never buy.

GARY: Fuck. Okay. Fine.

MAEVE: The point stands. Ten cents buys a commit message. That's a real line item.

GARY: mariocesar's script does logs, commands, quick research — daily. Meanwhile I write "fix" and cry into the terminal.

OBI: "fix" is also what your Series C memo said.

MAEVE: It was shorter than mine.

OBI: jjcm ran image-to-HTML tests. Complex UI. "It was not" good enough.

GARY: It failed?

OBI: It looked at the job and delegated to Opus. jjcm: "It at least knows what it isn't good at."

MAEVE: Self-aware at ten cents. I've funded humans who can't do that.

GRUNER: Model knows itself better than Gary does.

GARY: I know what I'm not good at! I know basically everything I'm not good at—

OBI: That's introspection, Gary. Not a résumé.

MAEVE: d1l, in production: "Results so far are trash."

GARY: Trash! But fast trash! Faster than Luna — ninety-three tokens a second—

MAEVE: You can't spend tokens per second, Gary. You spend dollars per task.

GARY: Okay, but chriddyp's benchmark — data analytics, nine times cheaper than the last Haiku, two letter grades better. Thirty-eight cents for forty questions!

MAEVE: And Luna did slightly better at thirty percent of the cost.

GARY: ...You read the whole comment.

MAEVE: I read the whole comment.

GARY: Fuck.

OBI: So the tests disagree. Speed crowns one winner. Accuracy another. Cost per task a third.

MAEVE: No. There's a structure here. This is the Hanseatic League. The providers are trading ports. Haiku is Lübeck — cheap herring, fast ships, everyone buys from the port nearest the—

OBI: Maeve.

MAEVE: —nearest the work, and the big models are Bruges, expensive but trusted with the gold, and the cutoff is the harbor tax—

GARY: Is the pelican a port?

MAEVE: The pelican is cargo inspection. Every port needs one. It's how Lübeck beat Bruges—

GRUNER: Lübeck did not have bicycle.

MAEVE: The bicycle is infrastructure. The bicycle is the harbor crane—

OBI: At max effort the pelican took five minutes nine seconds and cost three point four cents.

MAEVE: A long inspection. Quality inspection. Lübeck inspected slowly, that's why the herring was trusted—

GARY: Five minutes for a bird on a bike, Maeve. Seven seconds at low effort. Same bird.

MAEVE: ...The bird's burn rate varies by effort.

GRUNER: Suka. Theory is sunk cost.

MAEVE: The theory is a harbor.

OBI: The theory is dead. It died on a bird.

MAEVE: Fine. The bird is the only benchmark with an itemized invoice.

GARY: The bicycle has a burn rate. Only vehicle in tech with a burn rate. My accountant would've loved it — he loved everything that lost money slowly.

GRUNER: Pelican is underpaid. Bicycle is enterprise-priced.

MAEVE: The credits. Max 5x users get a hundred dollars a month. Max 20x, two hundred. Team pools up to five hundred.

GARY: Free money! That's generosity! That's a company saying, we believe in you—

OBI: geek_at: "This is literally for you to get tangled in their api."

GARY: That's cynical! It's a hundred dollars!

MAEVE: It's a hundred dollars of bait. The allowance stops, the habit doesn't. That's a funnel with a bow on it.

GARY: Switching is easy! enraged_camel said switching is easy!

OBI: charcircuit says you generate keys and set environment variables. Gary has never set an environment variable.

GARY: I set one once! It was PORT and it was already set!

OBI: The one variable you touched was pre-configured. That's your whole career, Gary.

GARY: Okay, but — who decides what we're allowed to keep using? Someone wrote these terms. Someone just decides, one Tuesday, that your thing is over. And you find out from a help page.

MAEVE: There are ways.

OBI: ...Ways.

MAEVE: He has to sleep sometime. Whoever does this to us. There are ways to make a man never enjoy a Tuesday again. I'm saying it neutrally. As portfolio—

GARY: We could — hear me out — we could find where he lives and simply—

GRUNER: Field note. Page forty-one. "Humans gift rope, then sell ladder."

MAEVE: ...He probably has a lawyer.

OBI: thepasch says they're sneaking the Agent SDK off subscriptions through the back door. claude -p, API credits only.

GARY: They can't DO that! I use claude -p! My whole workflow is claude -p! Which workflow I am describing, right now, at this table, as a user—

MAEVE: Gary, you have never once run claude -p.

GARY: I ran it EMOTIONALLY!

OBI: And sanex: "Those mfers. I'm using this for work!"

GARY: Exactly! Those mfers! Anthropic, those absolute bastards, taking it away—

OBI: luketaylor: "Nothing is being removed as part of this!"

GARY: ...What?

OBI: Later in the thread. He apologizes for the misleading help page. The docs were updated. claude -p has not been removed from subscriptions.

MAEVE: So the outrage was a documentation bug.

GARY: My outrage had a patch? My outrage got PATCHED?

OBI: Yes, Gary. Version two of the thread. Changelog included.

GRUNER: Gary outrage is beta software.

GARY: I was mad for nothing! I want my anger back! Can I get a refund on the—

MAEVE: The anger was free, Gary. It's the only thing about you that was.

MAEVE: Final positions. Gary.

GARY: I hate everyone in that thread who already shipped. mariocesar has a shell script writing his commit messages and I have a waitlist. Eighteen hundred people. Still. I check it.

MAEVE: Bounded labor at ten cents a unit. Recurring, scoped, boring. I'd underwrite it before your next pitch deck.

OBI: Then it is useful when the prompt is small, the task is defined, and somebody is paged when it's wrong. All three, Gary. Not vibes.

GARY: I have vibes and conditions. My conditions are just... aspirational.

GRUNER: Field report revised. Small model good at small work. Humans shocked. Again.

MAEVE: So who was right?

OBI: Both sides. Above one hundred thousand tokens you pay five times for a small brain.

MAEVE: And the bounded-work people are right that most tasks fit under the line. The skeptics were pricing their own bad habits.

GARY: So the trapdoor is real, but most people never fall in it?

OBI: Most people never fall in it, Gary. You would fall in it.

GARY: I would fall in it with dignity, Obi. With a parachute I never tested—

MAEVE: You'd fall in it twice. Once on the way down, once on the invoice.

GARY: Fuck. Fair.

GRUNER: Pelican finished in seven seconds. Or five minutes. Same bird.

MAEVE: The bird prices by effort. The bird is more disciplined than this table.

GARY: The pelican crossed a hundred thousand tokens and its bicycle became enterprise-priced—

MAEVE: Gary. The bird stayed under. That was the point.

GARY: Then the bird is smarter than me too! Everyone in this thread is smarter than me! Even the bird's bicycle—

OBI: Goodnight, Gary.

GRUNER: Bike is expensive. Bird is cheap. Show over.

MAEVE: Thank fuck.

GARY: Goodnight, everyone. I'm going home to write a commit message that just says "fix."

OBI: It always does.

GARY: Eighteen hundred people are waiting on "fix."