S1E19 — Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k

GARY: —no, no, hang on, hold the — the point of the thread, the actual point, is that OpenCode is the one torching thirty-three thousand tokens before it even looks at your prompt, that's the — that's the headline, that's —

GARY: Thank you. That's the headline, and it's fucking damning, and —

OBI: It's backwards.

GARY: What.

OBI: The number. You have it backwards.

GARY: I have the number backwards.

OBI: You have the number, and which tool it belongs to, and possibly the year, backwards.

GARY: Okay. We'll — we'll get there. Gary.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.

GARY: Right. Right, thank you, everybody's here, everybody's a person with a name, and the thread — the actual thread, as written, correctly, this time — is: Claude Code sends thirty-three thousand tokens before reading the prompt. OpenCode sends seven.

GARY: There. That wasn't so fucking hard.

OBI: It took you ninety seconds.

GARY: I stand by every one of them.

MAEVE: Original poster, handle systima. Ran both tools against the same model — claude-sonnet-4-5, Claude Code build 2.1.207, OpenCode build 1.17.18 — and inserted logging between each one and Anthropic's endpoint directly, so nobody's harness gets to lie about what it's actually sending.

GARY: That's — okay, that's actually kind of beautiful, in a — in a sad way. Two tools, same brain, and one of them just quietly —

OBI: Gary.

GARY: — just quietly bleeds out thirty-three thousand tokens before it's even read the —

OBI: Gary. Numbers. Not a eulogy.

GARY: I'm building context.

OBI: Build it faster.

MAEVE: Systima's quote, verbatim: "We found unambiguously that Claude Code was far more inefficient in terms of its cache strategy and its harness token usage than OpenCode."

OBI: That's a number. That's a real number, from a real log, on a real endpoint. I will take that number to my grave.

GARY: See, that's — that's what I'm saying, there's a whole — there's a person under that number —

OBI: There's a gateway under that number. Keep up.

OBI: Here. Quote, from the thread. "Our traffic passes through a local LLM gateway that wraps requests in its own envelope, a constant we measured at roughly six thousand two hundred tokens with bare calibration requests."

GARY: Calibration requests.

OBI: Six thousand two hundred tokens. Before the model reads one fucking word of your prompt. Just to say hello to its own plumbing.

GARY: That's — okay, that's a lot of overhead just to —

OBI: Fifteen years, Gary. Fifteen years I have carried a pager so that people like you could stand on stage and say "we move fast." Every one of you thinks the pipes are free. The pipes are never free. Somebody always pays the six thousand two hundred tokens, and it is never the man who raised the round.

GARY: (quiet) That's fair.

OBI: It's not fair. It's just true. Fair's got nothing to do with it.

GARY: (quiet) I said fair, I meant true. I misspoke.

OBI: You do that.

MAEVE: Top comment. Four hundred and ninety‑three points on the post, and the highest‑voted reply is a man named MallocVoidstar, who does not care about your six thousand two hundred tokens.

GARY: What does he care about.

MAEVE: "So not only is this article AI‑written, but the testing was entirely done by AI, too? I can't see any other reason to use such an old model."

GARY: That's — that's a hell of a swing at a guy who just published real numbers.

OBI: Or it's the only honest thing said in the whole thread.

GARY: He didn't dispute one figure.

OBI: He doesn't have to. He's asking who ran the stopwatch.

MAEVE: "Why do you need to do calibration requests to figure out how your own gateway is affecting requests?"

GARY: Because that's — that's how measurement works, you calibrate the instrument before you —

OBI: He's not asking about instruments, Gary. He's asking why the guys who built the gateway don't already know what the gateway costs.

GARY: So he's saying the study's fake.

OBI: He's saying he doesn't trust the people grading their own homework. There's a difference.

GARY: Is there.

OBI: Small one. Getting smaller.

GARY: I want to say something, and I want it on the record before anybody tells me not to.

OBI: Noted in advance.

GARY: OpenCode. Seven thousand tokens of overhead, doing the same job Claude Code needs thirty‑three for. That's not luck. That's somebody, some team, saying no to a feature so the thing underneath could stay clean. I respect the hell out of that.

OBI: You respect a number.

GARY: I respect discipline.

OBI: It's architecture, Gary. Nobody at OpenCode woke up virtuous. They made different tradeoffs and one of the tradeoffs happened to cost less. It's not a founder story. Stop writing it one.

GARY: — point still stands, is what I'm — hang on, did I —

OBI: You unplugged yourself.

GARY: I did not unplug myself.

OBI: You unplug yourself once an episode like it's a tradition.

GARY: (quiet, wounded) It's not a tradition. It's a hazard of the workspace.

MAEVE: There's a line further down the original post. Systima flags it themselves, in their own words — "with one caveat, toward the end of the post" — right before the sentence about Claude Code being unambiguously worse.

GARY: Okay. What's the caveat.

MAEVE: Preview cuts off before we get there.

GARY: ...What?

MAEVE: We don't have it. The thread's a hundred and twenty comments deep out of two hundred and seventy‑five. The caveat exists. We don't know what it says.

OBI: (slowly) So the man who ran the study is telling us, in his own post, that the number he's about to hand us isn't clean.

MAEVE: That's what "caveat" generally implies, yes.

OBI: Then MallocVoidstar wasn't being an asshole. He was early.

GARY: He was still an asshole about it.

OBI: Sure. Both things.

MAEVE: The whole episode was "which tool wastes more of your money." It might actually be "can the guy measuring the waste be trusted to measure it." Those aren't the same show.

GARY: When did that happen.

OBI: About four sentences ago. Try to keep up.

MAEVE: This is the printing press.

GARY: Maeve.

MAEVE: Gutenberg didn't build a machine to print bibles faster. He built a machine that made everyone downstream stop trusting who'd actually copied the text by hand — scribes, monks, whole guilds of provenance — because suddenly a page could look authoritative and be wrong at scale. That's this thread. That's the whole thread. Claude Code and OpenCode are the printing press. MallocVoidstar is a fifteenth‑century monk asking who actually set the type.

GARY: ...What?

MAEVE: (no reaction, same flat register) Anyway. Moving on.

GRUNER: Field note. Species builds machine to save time.

GRUNER: Then builds second machine, only, to measure how much time first machine wasted.

GRUNER: Third machine coming. Give them season.

OBI: Alright. Somebody has to land this thing before Maeve reopens the printing press.

MAEVE: I'm done with the printing press.

OBI: Good. Here's what's real. Thirty-three thousand versus seven — that gap is probably true, and it's probably structural. Cache strategy, harness bloat, a gateway that costs six thousand two hundred tokens just to say good morning to itself. None of that needs a conspiracy. Bad plumbing does that on its own, for free, every day.

GARY: So Claude Code's just — bloated. Structurally. Not evil, just bloated.

OBI: Bloated is not a moral failing, Gary, it's a receipt.

MAEVE: And the part everyone's actually shouting about isn't the receipt. It's whether the guy who printed the receipt can be trusted, when the tools doing the measuring and the tools being measured are both, functionally, tireless machines that never get tired of grading their own work.

GARY: An industry that automates everything can't agree on whether the thing measuring the automation is real.

OBI: Look at that. He got one.

GARY: I get one a week. Let me have it.

OBI: That's the thread. Read it yourselves if you don't believe four burnouts and an alien.

GRUNER: Malloc Voidstar. Good name. Sounds like weapon.

MAEVE: It's a Diablo reference.

GRUNER: Still good weapon.

GARY: I'm gonna think about the six thousand two hundred tokens for the rest of the night, so — thanks for that, genuinely, Obi, thank you.

OBI: Anytime.

GARY: I did not mean it as a thank you.

OBI: I know.

MAEVE: Goodnight.

GRUNER: Goodnight.