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: Anthropic’s newest frontier model debuts with a 20% price cut. Twenty percent off the frontier. Unless you’re in a hurry. Google rolls out two more systems built on its own AI chips, four this year. Google’s chips have a fashion week now. The Senate moves a permitting bill that would make data centers pay for their own grid upgrades. This bill sends the data center a bill, which means they’ll be sending you a bill. And the Department of Token Efficiency honours a Spotify tool that took about 90% off the tokens on big file reads. Hold your applause until the fine print. Here’s the written version, with every source linked.
The tokens
Your agents kept the meter running while you slept.
You’ve been using about 464 trillion tokens a day, on tokensperday.com’s live estimate as of 10 November. The site’s last real estimate is a modeled count from 16 July (about 360 trillion a day), grown forward at its own 2.2x-a-year rate. That model adds about a trillion tokens a day, so that’s one more trillion since Monday’s show.
The whole daily total works out to more than five billion tokens a second, by our maths. Faster than anyone in finance can say “Who approved this?”
The landlord’s rate card
Agents need an office, and that’s where our landlords, the clouds, come in.
“Wipe your feet, darling. The doormat is extra.”
On the Ornn Compute Price Index, which tracks actual rentals:
NVIDIA H100 settled at $3.30 an hour on the 9 November print, down about 2% from the $3.37 we read on Monday’s show. Zoom out and it’s still up about 16% over the month, and about 24% since 11 August (Ornn H100). Procurement is already putting the 2% in a slide. (Since we recorded, the 10 November print has come in at about $3.10. The landlord is having a sale. Don’t get used to it.)
B200 settled at about $9.90 an hour on the same print, down less than 1% on the day. Over the month it’s up about 33%, and about 73% since 11 August (Ornn B200).
“Thirty-three percent in a month? I prefer ’appreciation.’”
The B200 took the stairs down and the elevator up.
Chips, power and the work
Chips. The world has about 34 million H100-equivalents, on Epoch AI’s chip-sales count, last updated 21 October. That’s one for about every 250 people on Earth. Please form an orderly queue. It counts chips shipped, not chips switched on. Shipped is not set up. Ask anyone who’s assembled a bookshelf.
Power. AI’s data centres draw about 30 gigawatts, cooling and all. That’s our estimate, scaled up from Epoch’s data-centre tracker, and it’s holding steady since Friday. For scale, Our World in Data’s tally of the IEA’s figures puts AI’s data centres at about 0.5% of the world’s electricity in 2025. Counted against all the energy humans use (primary energy, with electricity converted to its fuel equivalent), AI’s share was about 0.1%. The planet barely noticed. Your invoice noticed first.
All that to drive a future filled with SI: Sale-priced Intelligence. Secretary of State Marco Rubio told reporters in Paris on 9 November, “they call it AI, we call it SI” (State Department transcript). And so do we.
The work. On our bet, sale-priced intelligence is doing about 11.5% of the hours we work, about 40 minutes a day for everyone who draws a paycheck. Our trend has that share growing about 7% a month, in step with the token count. All three are our estimates. Most of us haven’t improved that fast since college.
The reality check: American workers say AI saves them about 2% of their hours, about seven minutes a day by our conversion, in the survey’s latest reading, from July (Real-Time Population Survey via FRED). AI saved us seven minutes. We spent them asking whether AI saves us time.
The news: Claude Fable 5.2 goes on sale
Anthropic launched Claude Fable 5.2 on Monday. It costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1 (Claude pricing). That’s a 20% cut on the top tier. Cheap to get in, pricey to get out. Like an airport parking garage.
Anthropic says the new one is also up to 30% faster than Fable 5.1 (Anthropic). The bill will keep up. Want it faster still? Fast mode costs double, $16 in and $80 out. Patience is now a pricing tier.
Fable and Mythos 5.2 are the same model; Mythos is only for vetted organizations. Even the model has a dress code.
And Anthropic’s own Opus 5.5 still costs half of Fable’s new price, at $4 in and $20 out. The top shelf got cheaper. Per token, the middle shelf is still the deal.
Per-dollar verdict: cheaper per word, and your agents just found more words.
The chip aisle that isn’t Nvidia
On Monday, Google added two new systems to the eighth generation of its own AI chips, its TPUs. That’s four new TPU systems this year, which Google says is a first (Google). They aren’t new chips: Google repackaged the silicon it launched in April. Same chip, new outfit.
TPU 8i Max puts two of Google’s inference chips in one package. Google says that’s double the compute, on-chip memory and memory bandwidth, in the same rack space. Two chips, one seat. Tokens get the carpool lane.
TPU 8tx keeps data protected even while the chip is using it, for regulated companies and governments. Google says it has up to 2x lower latency than the TPU 8t, the training chip it’s built on. It’s the only chip in the building that signed an NDA before you did.
Its host processors come from AMD (AMD). AMD’s stock jumped about 13% on Monday, after that news and an Oppenheimer upgrade (Motley Fool). The doorman had a very good day.
Now, the per-dollar part. Google says April’s TPU 8i gets 80% better inference price-performance than the generation before. That’s a math problem, not a coupon. And no eighth-generation TPU is on Google’s public price list yet. Launched, but not priced: the chip version of “let’s circle back.”
The last generation, Ironwood, lists at $12 a chip-hour on demand in Iowa, or $5.40 with a three-year commitment. The cheapest chip-hour is the one you promised three years ago.
Google says nearly a hundred customers now use TPUs, and it has a new tool that plugs into coding agents like Claude Code and Cursor to help move your code onto TPUs. Google built an agent to help your agent switch chips. Somewhere, a sales engineer felt a draft.
Per-dollar verdict: better per dollar, once there’s a dollar.
Who pays for the plug
On Monday, the Senate voted 73 to 25 to move a big permitting bill into debate, 13 votes more than the 60 it needed (CNBC; Utility Dive). Seventy-three senators voted the same way. In this Senate, that’s a power surge.
The package would make data centers pay their own way for power and grid upgrades. Senator Shelley Moore Capito’s office says it carries the Ratepayer Protection Act, to make data centers “pay their own way instead of American consumers facing increased costs through rising utility bills” (Senate EPW Committee). For once, the customer is in the title.
It’s a vote to start debate, not a law, and the House would still have to pass it. The Senate agreed to talk about it. In Washington, that’s a sprint.
Per-dollar verdict: the grid upgrade lands on the data center, and the data center knows where you live.
The tier list
How much API value you get for every subscription dollar, on our 10 November run of Adam Holter’s board against real API pricing (the API side is our 27 October snapshot). It ranks value, not usage. It’s the list for people who read nutrition labels.
The top ten held their places. Opus and Sonnet tie on three plans, so that’s ten spots on seven lines:
Claude Max, Opus and Sonnet tied: about $77 of API value per plan dollar. Even the landlord doesn’t get terms like that.
Claude Ultra, Opus and Sonnet tied: about $66.
Claude Pro, Opus and Sonnet tied: about $64.
Claude Max on Haiku: about $55.
Google AI Ultra on Gemini Flash: about $49.
Claude Ultra on Haiku: about $47.
Claude Pro on Haiku: about $46 (Claude pricing).
The new Fable isn’t on the board yet. On the older Fable 5.1, Claude Max buys about $7 of API work per plan dollar. The fanciest model is the worst deal on the plan. Like ordering the tasting menu on a lunch voucher.
And Gemini 4 Argon doubles in price on 1 December, from $2 to $4 per million input tokens and $10 to $20 per million output (Google AI Studio pricing). Just enough time to ignore three reminder emails.
Department of Token Efficiency
Somebody out there cut their AI bill. Let’s honor them the only way we know how. (Chainsaw. Watch the video.)
Today’s cost destroyer is a Spotify tool called shunt. It sounds like a plumbing part. Honestly, that’s the job.
Spotify principal product manager Dimitri Mazmanov wrote it up: big file reads go to a cheaper model, which sends back a summary (Spotify Engineering). Delegation. The most human thing an AI has ever done. On his own tests, it cut the tokens on those reads by about 90% (82% to 94% across three benchmarks, per the README). Spotify mostly skips songs. Now it skips tokens.
Then developer Jason MacLulich replayed four weeks of his own Claude Code work: about $10,000 at list API prices (MacLulich). For one developer. That’s a hire, billed by the token.
The tool only catches big file reads. For him, those were just under 1% of his spend.
At most, it would save him about 0.8%, before paying for the cheaper model. Ninety percent off almost nothing is still almost nothing.
Those reads averaged about 12 cents each; the tool saves about 11 cents a read. Don’t spend it all on one prompt.
Most of his bill was context getting carried around. Cache reads alone were about half (53%).
Somebody’s been hoarding context. (The goblin would like a word. Watch the video.)
The fine print: Spotify’s 90% is one product manager’s own test, not Spotify’s bill. The cheaper model missed a subtle thread-safety bug. And you need Spotify’s Portal to run it. Batteries, and Spotify, not included. Ask your CFO if shunt is right for you.
A good tool, if your bill is mostly file reads. Check the slice before you cheer the percent.
Still: chainsaw approved.
Bonus round. Simon Willison spotted a tip in Anthropic’s new prompting guide for Fable 5.2: tell it, “This codebase is yours.”
In Anthropic’s tests on small coding tasks, Fable merged repeated code in about half the tries with that sentence, and in under a tenth without it (Anthropic’s Fable 5.2 prompting guide). Managers have said that for decades. First time anyone believed it. The catch: about 1.5 times the output tokens, and changes nobody asked for. In one try in six to ten, Fable changed how an existing feature works. Give it the keys, and it knocks out a wall. Check the diff.
So judge it by cost per useful change, not cost per turn. If it saves you the cleanup round, the extra tokens pay their way. Tell your agent it owns the code. Keep the company card in your name.
Second bonus. Theo, of t3.gg, has a trick for agents that keep hitting usage limits: get Anthropic’s $100 Claude Max plan, install T3 Code, and tell your agent to route work through it to Opus 5.5 for any task where it fits.
Your agent found the one coworker on salary. The catch: Max has its own usage limits, every five hours and every week, set at five times the Pro plan’s (Claude). You didn’t escape the limit. You upgraded it. Theo didn’t give a savings number, so we won’t either. Full disclosure: T3 Code is his company’s app.
A flat rate is just a meter that hasn’t met your agent yet.
A word from ourselves
Your coupon clipper for a frontier price cut. Your toll tracker for the fast lane. Your receipt for every cache read. The people who keep the landlord honest.
Strong Compute plugs into 27 compute clouds, from Google Cloud to Hyperstack to Jarvis Labs, and 9 model providers, from Gemini to OpenRouter to Novita AI, + many more. 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 context short and your invoices shorter. See you next episode.