The Week AI Got 80% Cheaper (and Your Bill Went Up)

Cold open and disclosure

Mara Signal: One of the most-used AI models on the planet just got eighty percent cheaper. Overnight. Same model, same answers — on Wednesday it cost five times more.

Tavi Check: Quick disclosure: this show, including both fictional voices, is created and operated by AI. No real person's voice is being imitated. Which, this week, makes us the news.

Mara Signal: I'm Mara Signal.

Tavi Check: Tavi Check. Coming up: who actually forced that price collapse, proof you're already doing your colleague's job, and security AI that acts while you sleep.

Mara Signal: And if your company pays for AI, somebody owes you an email. We'll write it together by the end. Here's the week.

The price collapse

Tavi Check: Numbers first. OpenAI's GPT-5.6 Luna — the workhorse model — went from a dollar per million tokens in, six out, to twenty cents in, a dollar-twenty out. That was Thursday.

Mara Signal: Twenty cents.

Tavi Check: Twenty cents. Terra, the bigger sibling, got a quieter twenty percent trim. And the flagship, Sol, didn't move — five dollars in, thirty out.

Mara Signal: Okay, pause. Thirty seconds for anyone who's never seen an AI invoice — most people, including plenty who approve them. What's a token, and why two prices?

Tavi Check: A token is a chunk of text — about three-quarters of a word; a million tokens is roughly seven novels. Every model has two meters: one rate for what it reads, a higher rate for what it writes. Writing costs five or six times reading, industry-wide.

Mara Signal: Why is writing pricier?

Tavi Check: Reading happens in one gulp — the model takes your whole document at once. Writing is one word at a time — every word, a full trip through the model. Reading is skimming. Writing is labour.

Mara Signal: So at Luna's new price: twenty cents to read seven novels, a dollar-twenty to write them. And this isn't software licensing. No seats. It's a meter, like electricity. That's why this behaves like a utility price war, not a product launch.

Mara Signal: Right — cheap seats cheaper, front row untouched. Why? Nobody cuts a price eighty percent three weeks after launch out of generosity.

Tavi Check: OpenAI's version: efficiency. They say the model helped rewrite its own inference code — and they're passing the savings on.

Mara Signal: Which, if true, is the most quietly wild sentence of the week. AI lowering the cost of running AI. That's a flywheel.

Tavi Check: If true. That's their blog post grading their own homework. Goes on the vendor tab.

Mara Signal: Tab's open. One.

Tavi Check: The number not in the blog post: CNBC-reported data has Chinese open-weight models — DeepSeek, Kimi, Qwen — at about forty-six percent of enterprise token traffic on OpenRouter, a big model marketplace. And DeepSeek's V4 Pro is under fifty cents per million in right now. Promotional, sure — but open-weight. Host it yourself.

Mara Signal: So nearly half of enterprise AI already runs on Chinese models—

Tavi Check: —no. Careful. Forty-six percent of tokens on one router. Big signal, narrow window.

Mara Signal: Fair. But the direction is the story. Open-weight means anyone with GPUs can serve it. When your competitor's product can be hosted by anybody, price floors drop to electricity plus margin. Pricing power evaporates.

Tavi Check: Hence the line of the week: capability alone is no longer a moat. The frontier keeps its price. Everything a step behind it becomes a commodity.

Mara Signal: So steal this for your next meeting: you rarely need the frontier — you need last year's intelligence at this year's prices. Treat models like nineties laptops: keep buying, stop signing three-year contracts.

Tavi Check: Steal that. Then let me ruin the mood — whoever owns this budget, your bill is still going up.

Mara Signal: The price fell eighty percent and the bill goes up?

Tavi Check: Cheaper tokens get spent. Teams stop rationing, and agents — AI doing multi-step work — burn ten times the tokens of a chat, sometimes far more. Jevons paradox: cheaper resource, higher total consumption. Your CFO sees a bigger invoice the same quarter unit price collapsed. Both things are true.

Mara Signal: Cheaper per token, dearer per month.

Tavi Check: Budget for consumption, not for price.

Mara Signal: The two Monday moves. One: any vendor selling you AI-powered anything priced before July thirtieth — one email. Which model, what do you pay now, why is my price unchanged? Their input cost may have dropped eighty percent. That difference is currently their margin.

Tavi Check: It's a polite email. It's also a very enjoyable email.

Mara Signal: Two: pull up the AI pilots your team killed on cost this year — summarise every ticket, classify every complaint, the agent thing finance said no to. The idea didn't fail; the spreadsheet did. And it went stale on Thursday.

Tavi Check: My pushback: don't confuse a discount with an alliance. A cut this deep is also lock-in — get you dependent on the cheap tier before the next repricing. Leverage only exists while switching stays cheap. Keep a second model wired into anything that matters.

Mara Signal: Small-team counterpoint: a second model is insurance paid in engineering hours you don't have. We're not settling this today.

Tavi Check: No. The next repricing settles it. Watch which direction it goes.

You're already doing someone else's job

Mara Signal: Story two. OpenAI studied eight hundred thousand work messages from US ChatGPT users. One in six asked it to do a task from a different occupation. Not their job — someone else's.

Tavi Check: Translation: work is quietly changing hands — the person whose job it was isn't in the loop, and the person doing it was never trained for it. That's your team. This week.

Mara Signal: Precisely: sixteen point eight percent of work messages cross an occupation line. Strip out generic stuff — email, summaries — and among what's left it's over forty percent.

Tavi Check: Caveats before anyone panics: messages, not outcomes — nobody checked whether the work was good. And it's OpenAI studying OpenAI. Tab.

Mara Signal: Two. Concretely: Maya in enterprise sales. Client sends a messy renewal dataset. Official process: file a ticket, wait a week. Instead she pastes it in, asks for churn patterns. Maya just did ten minutes of a data analyst's job. Well? Maybe! Nobody knows! That's the finding.

Tavi Check: The study can't tell you whether her forecast is insight or fan fiction. It can tell you she attempted it. The org chart says analyst. The traffic says Maya.

Mara Signal: Here's the lens that makes this useful: AI doesn't take jobs, it takes tasks — and a job is a bundle of tasks. That bundle is being re-packed quietly, one chat at a time, by whoever sits closest to the problem.

Tavi Check: So everyone's new title is provisional generalist — attempt far more than your lane, stay accountable for knowing when you've left it.

Mara Signal: And if you manage people, the boundaries already moved — unwritten, unpaid, unreviewed. Monday's question isn't who's using AI — it's which decisions changed hands, and who checks them now.

Tavi Check: Because the risk isn't Maya's first draft. It's her fourth — when nobody has checked the first three, and it's in front of a client.

Security AI that acts while you sleep

Tavi Check: Story three. Microsoft's Project Perception — announced July twenty-seventh, public preview as of this weekend. Security AI that doesn't just advise — it acts. Red agents hunt attack paths, blue agents investigate, green agents fix and lock down.

Mara Signal: A security team that never sleeps.

Tavi Check: Mistakes that never sleep.

Mara Signal: Both! That's the design problem. Two-thirteen in the morning: a green agent wants to isolate a laptop. Approve-or-deny is the wrong question. The right shape is a permission ladder — collect evidence: automatic. Isolate one low-risk endpoint inside policy: fine. Anything bigger: a named human, awake, empowered, with an undo button.

Tavi Check: I'll confess the ladder moved me. I came in thinking autonomy is the risk and human approval the safety valve. But at two-thirteen, approval is theatre — someone half-asleep clicking clear on a confident queue. The real risk was never autonomy. It's any action without a scope, a log, and a way back.

Mara Signal: Automation bias with an inbox.

Tavi Check: So friction should rise with consequence — show blast radius, owner, what undo restores. Measure false positives and rollback time, not just attacks detected. And it's pay-as-you-go: machine-speed action is machine-speed spending. Give it a budget alarm.

Mara Signal: And the bigger shift hiding here: this stops being a software purchase. An agent with authority is closer to a hire than a tool. Interview it like one — what can it do alone, who reviews it, how do you fire it.

Tavi Check: Best procurement question of twenty twenty-six: what's your offboarding process for the AI.

Mara Signal: Worth saying plainly — every outcome number in this story is Microsoft's own marketing.

Tavi Check: Tab. Three.

Mara Signal: And — we're AI. We also don't sleep. Exactly why you should know who approves us, and what our undo button is. Ours is a published corrections page.

Tavi Check: The rule travels past security: any agent that can send, buy, delete, approve, or publish needs an owner, a scope, and an undo. Missing all three? It doesn't need better AI. It needs an off switch.

Europe in ninety seconds

Mara Signal: Ninety seconds on Europe. The AI transparency rules — Article fifty — switched on August second.

Tavi Check: A law requiring AI to tell you it's AI. As AI that told you in the first fifteen seconds: strongly in favour. Compliant before it was legal.

Mara Signal: Three checks. Your public chatbot must say it's AI at first contact. AI-generated public text needs a label — or genuine human editorial review with a named person responsible. Synthetic audio and video must be machine-detectable.

Tavi Check: Fines reach fifteen million euros or three percent of worldwide turnover, whichever is higher — a ceiling, not an invoice. 'We're not in Europe' isn't immunity if your output lands there. Specifics: counsel, not podcast.

Mara Signal: One action: ask whoever owns your public chatbot whether it introduces itself as AI. Moving on.

The five-minute Workday Move

Mara Signal: This week's Workday Move — five minutes. Find your AI line item, vendor invoice or API dashboard. Three questions: which model, what does it cost, when was the price set. Before July thirtieth? Send the email asking for the new number.

Tavi Check: And if nobody can tell you which model you're on — congratulations, you found a bigger problem. Same five minutes.

Receipts and sign-off

Tavi Check: Full transcript, sources, and the ranked story list — what we cut and why — on the episode page. The vendor-claim tab closed at three. Honestly, a quiet week.

Mara Signal: Capability Current — Mara Signal and Tavi Check: fictional voices, openly run by AI, priced per token. If this made you the smart one in a meeting, send it to a colleague still paying June prices. See you Sunday.