In the first four hours after Fable came back, a developer named Nick Saraev spent over $1,400. In the next 24 hours, he spent roughly $1,000 more. He wasn't careless, and he wasn't rich enough not to care. He lit that money on fire on purpose — to answer a question a lot of us have been quietly dreading: now that the genuinely brilliant models are also genuinely expensive again, how do you keep using one without it bankrupting you?
What he found is worth your attention, because it flips the usual advice on its head. The way to cut an AI bill, it turns out, is almost never to buy a dumber model. It's to stop paying to ship the same junk to a smart one, over and over, all day long.
(A note before we start: every figure below is Nick's own measured or claimed result from his experiment, not an official guarantee. Your savings will vary by workload. The point is the pattern, not the exact percentage.)
The money wasn't going where he thought
Here's the thing that reframes everything. When Nick watched where his tokens actually went, the big line item wasn't the model sitting there "thinking hard." It was the plumbing — the invisible, repetitive traffic a coding agent generates just doing its job.
Take a single internal tool call. When Claude runs a command and reads the result, that result is often padded with the same information repeated a dozen times — "standard out, standard out, setting up fixtures, hydrating mocks" — none of which the model needs to see in full. And an agent fires hundreds, sometimes thousands, of these calls in one session. Nick ran a tool ("RTK," a token minifier) that strips each call down to only what matters. One call went from 612 lines to 4; from 36,700 characters to 177 — a 99% cut for the exact same information. Not every call is that wasteful, so his realistic estimate was a 30–50% reduction overall, with near-zero impact on quality. You're not making the model dumber. You're just no longer paying to say the same thing a thousand times.
Your prompts are carrying dead weight too
The same disease lives in your instructions. Nick pointed at a typical system prompt — the kind someone voice-notes in — that opens with "Hello, thank you so much for helping out with this project, we really appreciate all the work you do…" All of it is politeness with zero informational value to the model, and you pay for every token of it on every call.
Rewrite it for density and the meaning survives while the cost collapses. In his demo, a CLAUDE.md-style file went from 865 words to 211; the token count dropped from about 1,125 to 274. Do that across your prompts, memory files, and project context and, as he put it, you don't just save money — the model often gets sharper, because it's no longer wading through filler to find the actual instruction.
The other beats rhyme with this one. Don't let the agent read a 5,000-line log file as raw text — put it in a small database and let it query. Don't let it read a 20,000-line file start to finish to answer one question — have it sample the structure and jump to the 20 lines that matter. Even the language you prompt in shows up on the bill: in one test, the same simple query cost 51 tokens in English versus 118 in German. Every one of these is the same move — stop making the model chew through everything, and give it a way to fetch only what it needs.
The trap even the veteran fell into
Then came the confession that makes the whole video honest. Mid-experiment, Nick realized he'd been running about a dozen Chrome tool instances at once, each one silently loaded with its full context — quietly inflating his bill and, worse, confusing the model about which one to use. This happens to seasoned people. Context creeps in. A forgotten integration, a bloated set of skills, a pile of memory files, and suddenly a chunk of your budget is gone before you've typed a word.
His fix is almost funny in its simplicity: run a second agent on a machine you're not using, and have it check once a day whether anything crept into your context that wasn't there yesterday. Because — and this is the line that reframes the entire problem — "all this token management is context management." The bill isn't really about how smart the model is. It's about how much stuff you keep in the room with it.
The one trick that isn't free
To his credit, Nick doesn't pretend every tactic is a free lunch. One of them — forcing the model, via a strict rule, to only ever open a file when you explicitly point at it — genuinely trades quality for savings. Modern models are good at guessing where to look; clamp that instinct down and you'll save tokens, but you'll also have to be more precise or the model will miss things. It's a real tradeoff, and worth naming, because most "cut your costs" advice conveniently forgets that some cuts have a cost of their own.
The kicker: stop paying to over-think
The last one is the one I'd pin above your monitor. Modern models can "think" at different depths — feeding tokens through their own reasoning loop to get to a better answer. Left on adaptive, they tend to think far more than the task needs, and you pay for every one of those extra tokens.
Nick gave the same simple bug to a model on low effort and on its highest effort. Low: seven turns, about 1,028 tokens, 16 cents. Highest: nine turns, about 1,363 tokens, a few cents more — and the exact same bug found. In his testing the high setting routinely burned around 1.3–1.5x the tokens for no better result. His blunt aside: the vendor earns money on every token you spend, so "let the model decide how hard to think" quietly trends your bill upward. Default to low. Reach for high only when a task truly earns it.
What $1,486 actually bought
Strip away the specific tricks and the lesson underneath is the one that matters. Nick didn't save money by switching to a worse model. He saved it by refusing to overpay for a good one — by cutting the repeated tool output, the bloated prompts, the whole-file reads, the hidden context, and the reflexive over-thinking. Most teams staring at a scary AI bill assume the fix is a cheaper, dumber tool. The truer fix is almost always to look at how much waste you're shipping to the smart one.
That's the whole game as these models get more capable and more expensive at the same time: your cost isn't set by the model's price tag. It's set by the discipline of what you send it. Nick spent $1,486 to prove it. You get to skip the invoice.
Source: this article is based on Nick Saraev's experiment video "I spent $1,486 on Fable tokens so you don't have to" (YouTube). All spend figures, percentage savings, and token counts are his own measured or claimed results, not official numbers or guarantees, and will vary by workload. Model names and behaviors reflect the time of that video. Verify current model pricing and effort/thinking controls with each provider before relying on them.
