Listen to this episode on Spotify or watch it on YouTube.
Welcome back to The Cost of Intelligence, our comedy news show about the money behind AI. Each episode, Ben Sand reads the industry’s invoice and asks one question: are you getting more intelligence per dollar?
This week: Garry Tan ran Claude’s two big models through Claude Code on the same 52 coding tasks. The new Fable 5.2 cost 2.8 times as much per solved task as Opus, for scores too close to call. Same grades, pricier school. Nvidia’s August filing caps its guarantees on customer leases at $108.5 billion, most of it backing one Ohio campus leased to OpenAI. Nvidia used to sell the shovels. Now it’s the mine’s emergency contact. A former OpenAI researcher says six Chinese open models carry traces of Nvidia’s Blackwell chips, which are banned for export to China. And Elon Musk and other tech executives signed a petition asking Nvidia’s board to bring Jensen Huang back. Here’s the written version, with every source linked.
The tokens
AI reads and writes in tokens, chunks of words.
The world runs about 469 trillion tokens a day, on tokensperday.com’s live model as of 15 November (a July estimate, grown to today). That’s up about 2 trillion since Friday’s show. The tokens worked the weekend. Overtime’s on you.
On the site’s own path, that’s about 518 trillion a day by New Year’s Eve. The tokens already have New Year’s plans. Your budget wasn’t invited.
The landlord’s rate card
“A small weekend discount, darling. Don’t get used to it.”
On the Ornn Compute Price Index, which tracks actual rentals, as of the 14 November settle:
NVIDIA H100 settled at $3.19 an hour, down about 6% since Friday’s show and about 2% over the month (Ornn H100). The landlord gave back Friday’s raise. He kept the deposit.
B200 settled at $9.66 an hour, down about 1% since Friday, but up about 22% over the month (Ornn B200). The landlord left a mint on the pillow.
Chips, power and the work
Chips. Epoch counts about 34 million AI chips, measured in Nvidia H100s, its latest count, from 21 October (Epoch AI).
Power. AI data centres draw about 30 gigawatts. That’s our estimate, built on Epoch AI’s data-centre tracker. The International Energy Agency’s figure for last year puts AI at about half a percent of the world’s electricity, which is about a tenth of a percent of all the energy we use, our estimate (Our World in Data). Too small to notice. Big enough to need its own substation.
The work. Our name for it this week is SI: Seller-backed Intelligence. Intelligence where the store that sold you the couch also guarantees your apartment.
On our bet, SI does about 11.7% of the hours we work, about 41 minutes a day for everyone on a payroll. Our number ticked up a hair with the token count. Workers themselves say it saves them about 2% of their hours, per the St. Louis Fed’s July survey (FRED). That’s about seven minutes a day. SI just learned what middle management feels like.
These numbers are a work in progress. Got better data? Step boldly into our DMs.
The news: Nvidia’s filing
Nvidia’s August quarterly filing caps its guarantees on customer leases at $108.5 billion (Nvidia 10-Q). That’s more than all of Nvidia’s sales in that same quarter, about $96.2 billion (Nvidia CFO commentary).
Nvidia only pays if a customer stops paying its lease.
Most of it, $105 billion, backs one campus in Ohio, leased to OpenAI.
The other $3.5 billion covers other AI clouds.
Nvidia sells you the chips, then vouches for your rent. Every parent of a college kid knows that deal.
The Ohio campus is planned at about 4.25 gigawatts. It goes up in nine phases, and the first isn’t due before 2028. Nine phases. Even the data centre is on a payment plan.
Per-dollar verdict: the cheapest chip is the one somebody else is on the hook for.
Chips in China
Will DePue used to be a researcher at OpenAI. He says he found traces of Nvidia’s Blackwell chips inside six Chinese open models: the way their weights were set up matches Blackwell, not the older Hopper chips (The Decoder; TechCrunch; The Register).
The list includes DeepSeek’s V4 models and Kimi K3. He says it shows they had access to the chips. He can’t say how they got them, or where the training ran. They may be free weights, but they’ve been training at an Equinox gym.
They’re not just fit, they’re popular too. This week, Chinese models carried more than eight in ten tokens on OpenRouter’s top ten list (OpenRouter rankings).
Per-dollar verdict: the cheapest tokens may depend on the priciest chips.
Nvidia’s boardroom
Nvidia announced Jensen Huang’s resignation on Thursday, and now there’s a push to bring him back.
Elon Musk and other tech executives signed a petition asking the board to reinstate him (TechCrunch). The petition says it has more than 100,000 signatures. Ben can’t get four people to sign a birthday card.
Investors Brad Gerstner of Altimeter and Gavin Baker of Atreides wrote to the board too, asking it to keep the door open to his return (CNBC).
Nvidia reports earnings on Tuesday, 17 November. The board has a petition, an investor letter and an earnings date. Its inbox needs an inbox.
Per-dollar verdict: 100,000 signatures, zero discounts.
The tier list
How many dollars of API usage you get per dollar of subscription, on our 15 November run of Adam Holter’s method with Real API Pricing. And they’re off!
Claude Max breaks first, at about $77 of API value per plan dollar.
Claude Ultra runs second at about $66, and Claude Pro is a neck behind in third at about $64. Claude takes win, place and show. Somebody check who owns the track.
Then a gap, and into the turn comes the pack: Claude Max on Haiku about $55, Google AI Ultra on Flash about $49, Claude Ultra on Haiku about $47, Claude Pro on Haiku about $46, and X Ultra at about $44, its floor.
Ninth and tenth: ChatGPT’s two Pro plans, about $28 and $27.
Trailing the field, last of 44: Google AI Pro on Gemini Argon, at about $1.40. Argon hasn’t even left the paddock.
On 1 December, Argon’s API price doubles (Gemini API pricing). Same model, double the bill. That’s not an upgrade. That’s a cover charge.
This whole board has held steady since Friday. Nobody’s passing anybody. The finish is a photocopy.
Department of Token Efficiency
Three cuts today. We’ve sharpened the chain.
Cut one: Garry Tan, the head of Y Combinator. He ran two of Claude’s big models through Claude Code on the same 52 coding tasks (Garry Tan on X; g-harness-evals).
The new model, Fable 5.2, solved 83% of the tasks. Opus did a bit better at 85%. Too close to call.
But per solved task, Fable cost $1.83 and Opus only $0.65. That’s 2.8 times as much.
Tan’s advice: keep Opus as your coding default. Same grades, pricier school.
One poster on X said Fable ate their whole weekly limit in 20 minutes (PrecyartG on X). Brunch lasts longer.
Then Tan kept the model and swapped the harness. On Opus, OpenClaw cost 40 cents per solved task and Claude Code 21 cents. Nearly double, on the same model. Tan won’t crown a best harness. Swap the harness, keep the brain: half the bill goes down the drain.
The fine print: Tan’s own test. 52 tasks for the models, 20 for the harness. Y Combinator has a stake in two of the harnesses, and two add-ons he tested are his own. Side effects may include sudden loyalty to the model you already have.
Cut two: Deepnote, which makes data notebooks. It priced a month of an AI agent watching one company’s Slack (Deepnote).
On Fable, it came to about $4,700 a month.
OpenAI’s Sol did it for $281. That’s 94% less, and Sol finished second on Deepnote’s scoreboard.
Agents that butt in cost 1.8 times as much as agents that wait their turn. Finally, a benchmark for the guy who talks in every stand-up.
The fine print: Deepnote’s own benchmark, 50 workplace scenarios. The monthly bill is Deepnote’s extrapolation, not an invoice. Ask your manager if reply-all is right for you.
Cut three: HeyGen, which makes AI video. It built its own harness for video work, called HyperFrames, and ran 17 video tasks on the same model (HeyGen).
HyperFrames cost 30% less than Claude Code.
It needed 35% fewer model turns.
A judge model rated the results even.
Even Claude Code pays a generalist tax. The specialist gets a refund.
The fine print: a vendor grading its own product. Percentages only, no dollar figure. 17 tasks, judged by a model, not by people. Results not typical. Your video may vary.
Still: chainsaw approved.
A word from ourselves
Your credit check on the chipmaker’s guarantee. Your fingerprint kit for the model you’ve been running. Your receipt for every task solved. The people who keep the landlord honest.
Strong Compute plugs into 27 compute clouds, from Azure to Hyperbolic to Salad, and 10 AI vendors, from Anthropic to OpenRouter to Novita AI. Bring your own keys and see exactly where the GPU hours and tokens go.
Because somebody has to read the bill.
Try it at strongcompute.com
Keep your contexts short and your invoices shorter. See you tomorrow.










