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: AI-tool tester aifindr gave four Claude models the same prompt, and the bills ran from 9 cents to $4.97. On TPD Tuesday, ten dollars buys 286 times as many tokens on one host as on another. Jensen Huang says he would return as Nvidia’s CEO if the board asks. Alibaba says Qwen 4 Max cuts the cost per task by about 60%, on its own test. And Bloomberg reports that orders for Anthropic’s IPO are close to seven times the shares on offer. Here’s the written version, with every source linked.
TPD Tuesday: what $10 of tokens buys
TPD is tokens per dollar: how much model your money buys. We took $10 of output tokens at list price and counted. Our arithmetic: $10 divided by each output price per million tokens. Output tokens only, no caching, batch or promo prices.
DeepSeek, on its own site, current Flash model: 16.7 million tokens off-peak ($0.60 per million), 8.3 million at peak ($1.20) (DeepSeek pricing).
Relace, a host on OpenRouter running DeepSeek’s older V4 Flash in a four-bit version: 71 million tokens ($0.14 per million) (OpenRouter). Seventy-one million tokens, for the price of lunch.
Claude Haiku 5.5: 20 million, for prompts under 100,000 tokens ($0.50 per million) (Claude pricing).
Qwen 4 Max, Alibaba Cloud’s international price: 2.2 million ($4.50) (Qwen Cloud pricing).
Gemini 4 Hydrogen: 1.3 million ($7.50, thinking included) (Gemini API pricing).
Claude Opus 5.5: half a million ($20). Claude Fable 5.2: a quarter of a million ($40) (Claude pricing).
Same ten dollars. Now you’re rationing. On the Relace host, $10 buys about 286 times as many tokens as on Fable.
Tokens per dollar isn’t work per dollar, though. A model that thinks longer spends more tokens on the same job. So set it next to what a flat plan buys.
Tier List Tuesday
How many dollars of API usage you get per dollar of subscription, on our 17 November run of Adam Holter’s method with Real API Pricing. And they’re off!
Claude Max on Opus leads wire to wire, at about $77 of API value per plan dollar.
Claude Ultra runs second at about $66, and Claude Pro is third at about $64.
Into the turn: Claude Max on Haiku is fourth (about $55), Google’s AI Ultra on Gemini Flash holds fifth (about $49), Claude Ultra on Haiku is sixth (about $47), Claude Pro on Haiku seventh (about $46), and X Ultra eighth ($43.70 or better).
Ninth and tenth: ChatGPT’s $200 and $100 Pro plans, both in the high twenties. Last of the top ten. Nice silks, though.
That’s API dollars. Count tasks per dollar instead, and OpenAI’s Luna on the Plus plan comes out ahead, by our estimate. Claude wins the race. Luna eats the least hay.
The news: Nvidia
Nvidia announced Jensen Huang’s resignation as CEO last Thursday, 12 November. Colette Kress is interim CEO, and the board is searching for a permanent one (NVIDIA).
On Tuesday, Huang posted a statement on X (Jensen Huang on X). He writes that he “did not live up to the standards expected of everyone at NVIDIA” and that he is committed to cooperating with the board’s independent investigation. He also writes that if the board decides his return is in the best interest of the company, he “would be honored to answer these calls and return as CEO.”
NBC reports that Nvidia has hired an outside law firm to investigate, that Nvidia shares fell more than 4% on Monday, and that President Trump says he would not mind if Huang came back (NBC News).
Nvidia reports its third-quarter results after the market close on Tuesday, 17 November, with a call at 5pm Eastern (NVIDIA). The results came after we recorded. We’ll open the envelope next episode.
Per-dollar verdict: nothing in the statement changes what a chip hour costs. The results might.
Alibaba’s Qwen 4 Max
Alibaba launched Qwen 4 Max on Monday. It says the model cuts the average cost per task by around 60% compared with its last Max model, and uses 35% fewer tokens on its high-effort setting (Alibaba Cloud). Alibaba reports the two numbers side by side and doesn’t say how they relate.
Those results come from Alibaba’s own benchmark, Work360. Alibaba wrote the test and graded it too. That’s how most of us do our self-reviews.
Per-dollar verdict: sixty percent off, once someone else grades it.
Anthropic’s IPO
Bloomberg reports that Anthropic’s IPO order book closed on Tuesday, 17 November, and that orders are close to seven times the shares on offer. Pricing is set for Wednesday, and trading is due to start Thursday on the Nasdaq under ANT (Anthropic).
A hot book can push the price up. For investors, that means paying more for the same slice. Nearly seven to one: that’s a concert presale with a ticker.
Per-dollar verdict: the more they want it, the less your dollar buys.
Department of Token Efficiency
Chainsaw fueled. Spreadsheets open.
Cut one: DigitalOcean tuned how it runs Meta’s Llama 3.3 70B (DigitalOcean).
Cost per million tokens fell from $5.80 to $1.47, about 75% cheaper.
It ran on two H100s instead of four. Half the H100s, same model. Every reorg promises that. This one delivered.
The fine print: DigitalOcean’s own test, against an untuned setup. It’s a GPU-utilization cost, not a price anyone charges. The gains fade under heavy load (above about 16 concurrent requests).
Cut two: Octomind tested its coding agent against OpenAI’s Codex (Octomind).
On a typical task, Octomind used 25% fewer tokens, and its cost per task was 13% lower.
Same score, fewer tokens. The kid who skipped the reading and still passed.
The fine print: the maker ran the test. Both agents passed 38 of 65 one-shot runs.
Cut three: aifindr tests AI tools on X. It gave four Claude models the same prompt to build the same throwaway prototype (aifindr on X).
Claude Haiku 5.5 did it for 9 cents.
Claude Fable 5.2’s whole bill came to $4.97.
For a throwaway, the tester picked Haiku.
Five dollars for a prototype you’ll delete by Friday. That’s a very expensive napkin.
The fine print: one prompt, one run each, posted by a small account. Not a benchmark.
Ask your CFO if token efficiency is right for you.
A word from ourselves
Your seatbelt for earnings season. Your coupon clipper for the sixty percent off. Your second opinion on every rate card. The people who keep the landlord honest.
Strong Compute plugs into 27 compute clouds, from Azure to Crusoe to Hyperstack, and 10 AI vendors, from OpenAI to Fireworks AI to Simplismart. 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.



