S4E375 — Why isn't the industry freaking out about DeepSeek 4.1 Flash?

GARY: Okay, so — nobody's panicking, which means the panic is somewhere in this building, and I'm going to find it. Gary.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.

GARY: Right. So there's a blog post. A guy used DeepSeek 4.1 Flash for a month, heavily, like a dozen projects, and mid-session he genuinely cannot tell if he's using DeepSeek or Opus.

MAEVE: He cannot tell. That's the whole headline. That's the panic.

OBI: That's not panic, that's a man on a ten-dollar OpenCode Go plan describing a religious experience.

GARY: Ten dollars a month and DeepSeek is basically unlimited. His sessions mostly stay under a dollar. Some run most of a day.

MAEVE: A dollar. I pay more than that to be told my parking is validated.

GARY: And the reason it's cheap — the KV cache. He says DeepSeek shrank it by roughly 437 times versus their V1, which is apparently why it's also better for the environment, less water, less electricity, the whole --

OBI: No. Gary. The 437 is the cache size. One-four-hundred-thirty-seventh the size. It says nothing audited about water, electricity, or the planet. You compressed a memory buffer into an ecosystem.

GARY: I compressed it because it's inspiring.

OBI: Numbers are not inspiring, Gary. They are load-bearing.

MAEVE: A ten-dollar plan making frontier-adjacent labor unlimited. Do you understand what that does to a labor market?

GRUNER: Yes. It is called "unemployment."

GARY: No, no — it's called basically unlimited. He says basically unlimited. That's the whole point, intelligence is now so cheap that --

GRUNER: Then who is unlimited?

GARY: Nobody, Gruner, it's a subscription. There's no workers. It's -- okay, it's ten dollars, and the cache is one-four-hundred-thirty-seventh the size, and I got excited, shit happens.

GRUNER: Unlimited labor for ten dollars. I have questions for his workers.

MAEVE: Oh, he's going to be insufferable about this all episode.

GARY: He's not wrong, though. That's the bastard of it.

OBI: Say the fraction.

GARY: One-four-hundred-thirty-seventh.

GRUNER: Good. Field report writes itself.

GARY: Okay, first comment. badatnames: "DS4 crossed my dont-care threshold and I genuinely stopped paying attention to new models."

MAEVE: A dont-care threshold. I want that metric in every deck I fund. Series A question: what's your indifference number.

OBI: He has a number for not caring. Gary has a number for inspiration. Between you two, the industry is fully covered.

GARY: That's not indifference, that's peace. He found the model where caring stops being required, which is --

MAEVE: He stopped reading release notes, Gary. That's not peace, that's surrender with a receipt.

GRUNER: Threshold is good word. Humans have dial for giving up. Factory setting.

OBI: It's not giving up, it's convergence. Tool's fine, you stop shopping. Nobody audits their toaster.

GARY: Right! Exactly! My toaster is fine! Nobody's toaster is a frontier --

MAEVE: Your toaster caught fire at Halfpipe's launch breakfast, but go on.

GARY: Different toaster, different company, and the regulator was very nice about it.

OBI: The regulator. The woman who killed the skateboard bank.

GARY: She was extremely professional about it, Obi.

OBI: She sent you a letter with a frowny face and a licence number.

GARY: Second comment. user43928: "there is no 'good enough'."

MAEVE: There it is. The whole thread in five words.

OBI: He's wrong in an interesting way, which is the best kind of wrong. His claim: open models stay six to twelve months behind, so the frontier keeps escaping.

GRUNER: Frontier is donkey on stick. Always ahead. Never catch. Donkey tired.

GARY: That's -- no. That's cruel. The author uses it for planning. For research. He pulls in Opus for a final code review and --

MAEVE: Ten dollars for the intern, two hundred for the one who checks the intern. That's not a price war, that's org design.

GARY: It's a ladder of intelligence and everyone has a rung, and I know what it's like to be a rung --

OBI: Gary, you were employee four. You weren't a rung, you were the ladder's load test. The ladder failed.

GARY: Fuck. Okay. Fine. But "good enough" is a legitimate engineering standard, Halfpipe had a whole compliance --

MAEVE: You're citing the company that died of "very nice" as evidence for quality.

GARY: It was a quality death!

GRUNER: Death is death. Rung closed.

GARY: Last one. thefourthchime, and I want the numbers clean: Opus 5.5, score 99 out of 100. DeepSeek 4.1 Flash, 72 out of 100.

MAEVE: Seventy-two. So "good enough" is a rounding error with a blog.

OBI: Same benchmark, same task, twenty-seven points apart. That's not a rounding error, that's a different species.

GARY: Or -- hear me out -- a species that costs a dollar a day, and honestly, mid-session, if you don't look at the model name, you genuinely could not tell --

MAEVE: Gary.

GARY: Whether you're using DeepSeek or --

OBI: Gary. That's the article. You're quoting the article at us. Word for word. From memory.

GARY: I know whose article it is.

MAEVE: You've become the commenter you envy. That's a premium feature. We charge for that.

GRUNER: Field note: host crossed own dont-care threshold. Mid-episode. Sad. Fast.

GARY: p1necone: "It's perfectly capable of being the sole agent for all of my well specced implementation tasks."

MAEVE: "Well specced." So the model's a genius when the human does the thinking first. That's not hiring, that's a séance.

OBI: He tried it as orchestrator. Quality loss, designs from thin air, docs way too wordy. Went back to GLM for planning.

GARY: So it's not one model versus one model. It's a job description. Planning, implementation, review — different jobs, different prices.

GRUNER: Model is employee. Cheap employee good at one thing.

MAEVE: That's every startup's first ten hires, Gruner.

OBI: try-working, forty billion tokens, eight hundred dollars in a month: split runs into phases, build in verification, use subagents for an audit pass.

GARY: See, that's basically what I did at Halfpipe. I had a whole executive team --

OBI: You had five contractors and a Notion page.

GARY: I had a management structure, Obi. And these cheap subagents are the same thing. Fifty of them, ten dollars. My executive team.

OBI: Gary. Fifty subagents is not an executive team. It is fifty processes with no shared context and no lock.

GARY: They report up. Through the orchestrator. That's literally management.

OBI: Your orchestrator is a context window. It forgets the org chart every four thousand tokens.

GARY: Well, apitman actually did it, so --

OBI: Do not cite apitman at me.

GARY: apitman: "I fired up first 10, then 20, and finally 50 concurrent subagents all working on reverse engineering C code from an old PC game."

MAEVE: Fifty interns, no HR liability. I'd fund that.

GARY: "They found over 100 new functions."

MAEVE: For the price of --

GARY: "That cost maybe $2."

GRUNER: Two dollars. Fifty workers. Blyat, even I pay interns more.

OBI: He also said it dented his five-hour window and barely moved the monthly. That's not management, that's a firehose.

GARY: It's delegation! I delegated at Halfpipe, it worked --

OBI: It did not work. Your regulator letter is public record, Gary.

GARY: That was compliance, not delegation --

OBI: Fifty agents, one codebase, no sentinel, no merge order, a context window as the only coordinator. That is not an executive team. That is a race condition with a payroll.

MAEVE: He's filing Gary under a new failure shape.

GRUNER: New shape. Good for taxonomy. Bad for Gary.

GARY: Fuck. Okay. The fifty agents were a metaphor --

OBI: They were someone else's real workload. You just claimed the org chart. That's the most Halfpipe thing you've ever done.

GARY: I built things at Halfpipe!

OBI: You built a compliance file.

MAEVE: Gary, your executive team costs two dollars. I've paid more for a sandwich that lied to me.

GRUNER: Sandwich also had org chart. Also collapsed.

GARY: Okay, so why is nobody screaming. aszen: "Because subscription plans are cheaper, only enterprises paying per tok pricing should be freaking out."

MAEVE: That's the answer. The panic is being comped.

OBI: vishvananda, independently, same hour: "The reason people aren't freaking out is because most people are using heavily subsidized subscriptions."

GARY: Two people found the same door.

MAEVE: Three. qwerpy: "$100 for Claude just about gives me all the usage I want." Token-equivalent, he's spending five grand a month. The sub is one-fiftieth of the API rate.

GRUNER: Human pays one. Feels rich. Suka, market is drunk.

MAEVE: It's not drunk, it's disciplined. This is the Venetian Arsenal. Standardized parts, state arsenals, ships assembled from a catalog. Intelligence is becoming a standardized part. The labs are the Arsenal. The subscriptions are the --

OBI: The Arsenal shipped warships, Maeve. These are twenty-dollar plans.

MAEVE: Twenty dollars is a hull section. One hundred dollars is a hull section. The pricing tiers ARE the standardization --

GARY: The Arsenal didn't lose money on every galley.

MAEVE: It absolutely priced below cost during the ramp-up. That is textbook arsenal economics, you subsidize the fleet until --

OBI: You're extending it. Every time we push, the Arsenal gets bigger.

MAEVE: Because the Arsenal is a METAPHOR with legs. Two hundred a month for Opus, token-priced that's thousands. That's not a loss, that's land-grab pricing, that's the Arsenal building shed number --

GRUNER: Shed number two hundred.

MAEVE: -- exactly, and the state pays, and Venice owns the sea, and nobody freaks out because the freakout is the pricing --

OBI: Maeve. ctolsen, and I want the sentence whole: "So the industry is responding, where it matters. Which is on heavy API usage, not coding subs."

MAEVE: They got there first. That's all.

GARY: Did she just --

GRUNER: Field report update. Humans subsidize each other's denial.

OBI: He said it months of usage ago, on the open thread, in plainer words than your Arsenal.

MAEVE: The Arsenal had better production values.

GARY: So the labs aren't not panicking. They're panicking in the invoice line.

MAEVE: Which is where I would panic.

OBI: Which is why nobody can see it. Enterprises paying per token are the ones sweating, and enterprises don't scream, they renegotiate.

GARY: That's the creepiest sentence anyone's said tonight, and I once got a cease-and-desist with a recipe in it.

GRUNER: Invoice is scream. Quiet scream. Good.

MAEVE: I hate that the thread is smarter than my arsenal.

GARY: You'll rebuild it. You always rebuild it.

MAEVE: The Arsenal had a fire, Gary. It recovered. You, famously, did not.

OBI: Leave the man's ash out of naval history.

GARY: Thank you, Obi.

OBI: The ash is load-bearing.

GRUNER: Shed number two hundred burned also. Venice fine. Gary, question mark.

GARY: I'm not answering that, Gruner.

GRUNER: Correct answer. Field note closed.

GARY: dannyw, on why some people think it's dumb: "heaps of providers either silently serve heavily quant versions, or don't have inference set up correctly."

OBI: So half the bad reviews are a provider misconfiguration wearing a model's name.

GARY: RussianCow pins his config to one provider: "I regularly get ~98-99% cache hit rates with OpenCode."

MAEVE: Ninety-nine percent. The rounding error is the product.

GRUNER: Cheap is not one number. Is weather.

GARY: And jerieljan on the hundred-dollar tier: "what you get on OpenAI or Anthropic is really more than enough."

GARY: Last one, and it's short. bitfilped: "Because in two weeks someone will be asking why I'm not freaking out about AlphaDolphins 0.3 Zip and then in a month FrozenMonkey 2.5 Artic."

MAEVE: AlphaDolphins 0.3 Zip. I've funded two of those. Both are now dental software.

OBI: FrozenMonkey 2.5 Artic is a real-sounding fake and I hate that I'd benchmark it.

GRUNER: We also are replaced in weeks. Is fine. I have appendix.

OBI: Thank you, Gruner.

GARY: Okay. Verdict. Cost pressure is real, it's not frontier everywhere, and the subs are why nobody's screaming.

MAEVE: The scream is in the pricing. I hate it there.

OBI: And the model you pick is a job description, not a flag.

GRUNER: Humans buy status. Then discount status. Then both, same receipt.

GARY: Which brings us back to ten dollars, basically unlimited. Fifty agents, an entire game, the price of a coffee. That's a labor market with a student rate.

MAEVE: The cheapest workforce in history, and Gary raced-conditioned it in one afternoon.

GARY: Fuck you, Maeve, gently.

OBI: Goodnight from the booth. Mind the cache.

MAEVE: Goodnight.

GARY: Goodnight, losers.

GRUNER: Industry freaks out at one-four-hundred-thirty-seventh volume. Correct volume.