Dev.to AI 🤖 Ai 👁 0 📖 5 min read

Your AI Bill Has a Second Price Dial — and It Can Out-Swing Model Choice

Everyone prices the model. Nobody prices the dial. When a business compares AI costs, the conversation is always about model choice: Fable versus Opus, the $8-per-million-token model versus the $4 one, the premium lane

Everyone prices the model. Nobody prices the dial.

When a business compares AI costs, the conversation is always about model choice: Fable versus Opus, the $8-per-million-token model versus the $4 one, the premium lane versus the cheap lane. That's one price dial. But there's a second one sitting right next to it — the reasoning-effort setting, the value that tells the model how long to think before it answers — and fresh hands-on testing suggests it can swing your bill harder than the model choice itself.

What a controlled test measured

The team behind the open-source Atomic Agent coding tool ran a controlled comparison in mid-November 2026, right after Claude Fable 5.2's launch. One real task: a roughly 1,400-line Python library with 13 planted bugs and one missing feature, scored against 59 hidden tests the agent never sees. Each setting ran three times. The averaged results:

Setup Hidden tests Time Cost per task
Opus 5.5, medium (its API default) 59/59 4.2 min $0.97
Opus 5.5, high 59/59 9.6 min $1.70
Fable 5.2, medium 59/59 4.2 min $1.87
Fable 5.2, high (its API default) 59/59 26.6 min $7.70

Every single run passed all 59 tests. The most expensive setting cost about eight times the cheapest one — for identical results.

The effort dial out-swings the model dial

Two comparisons fall out of that table, and they're the whole point:

  • Swing the model dial (Opus 5.5 → Fable 5.2, both at medium effort): the bill went from $0.97 to $1.87 — 1.9x, almost exactly the 2x gap in the models' list prices.
  • Swing the effort dial (Fable 5.2 at medium → Fable 5.2 at high, same model): the bill went from $1.87 to $7.70 — 4.1x.

On this task, the effort setting moved the bill about twice as much as the model choice did.

The New Stack ran its own comparison and found the same shape: extra reasoning "did not raise Fable's score on any of the three versions" of its test, while the costlier setup ran an estimated $3.29 versus $0.99 for runs that matched the scores. And coverage of Artificial Analysis's effort-level data reports Fable 5.2 at low effort coming out cheaper, faster, and scoring higher than Opus 5.5 at high effort — the dial swamping the model from both directions.

Where the dial hides in your stack

The trap isn't that effort exists. It's where the default sits.

Fable 5.2's default on the Claude API is high effort. Opus 5.5's is medium. In the Atomic Agent runs, simply swapping the model ID — changing nothing else, sending no effort value — made the same task 7.9x more expensive and 6.3x slower, because the swap silently moved the default from medium to high. In their words, that one change raised the bill more than anything else they tested.

For a small business, this dial hides in three places:

  1. Scripts and integrations that call an AI API directly. If your Zapier step, custom script, or automation flow doesn't set a reasoning-effort value, you bought the model's default — and on premium models that default is often the expensive one. Dropping Fable 5.2 from high to medium in the test cut cost 76% and time 84% with identical results.
  2. Third-party AI tools and agents. Whatever effort setting the vendor baked in is on your invoice. If a tool can't tell you what effort level it runs at, treat that answer — or the absence of one — as vendor-evaluation data.
  3. "Extended thinking" and "deep research" toggles in subscription tools. Different name, same dial: you're choosing how long the model thinks, and thinking is billed by the token.

The three-attempt dollar benchmark

Before paying premium effort on any recurring AI job, price the dial the same way the testers did:

  1. Pick one recurring task you already run weekly — a report, a content draft, a ticket summary.
  2. Run it three times at the default effort setting. Record the dollars per run and whether it finished to your standard.
  3. Run it three times at high effort. Same recording.
  4. Compare dollars per finished task, not per attempt. If high effort didn't finish anything the default didn't, the premium bought nothing — you just paid for thinking.
  5. Write the winning effort level next to the job, and re-run the benchmark quarterly. Defaults and pricing move with every release; your note is the audit.

Define "finished" before you run, not after. "Looks plausible" is not a finish. Use a checkable standard: tests pass, the fact-check list comes back clean, the draft contains the three required sections.

When high effort is worth it

None of this says high effort is a rip-off. The Atomic Agent team is explicit about the limit of their own test: on a task hard enough that the cheaper lane fails, the extra spend could easily be worth it. Anthropic's own guidance follows the same shape — start at high, step down where your results hold, and measure before paying for anything above it.

The split that fits most small businesses:

  • Effort up: novel, high-stakes, or genuinely hard work — the first version of something, work where a wrong answer costs real money.
  • Effort down: routine, repeated, verifiable work — the weekly report, the standard draft, the recurring classification. If you can check the output in two minutes, the model doesn't need to think for twenty-six.

Honest boundaries

  • The headline test is one task at three runs per setting, and every setting scored 59/59 — it measures spend-to-reach-the-same-result, not which model is smarter.
  • Defaults differ by surface. Fable 5.2's API default is high, but Anthropic made medium the default in its own apps and Claude Code. Price the dial where you actually use it.
  • The test didn't cover the xhigh or max levels on either model — the 76% cut came from stepping high down to medium.
  • Effort names, defaults, and prices move fast. The durable habit is the benchmark, not memorizing this month's settings.

The audit nobody runs

Most businesses that pay for AI cannot say which of their recurring jobs run at which effort level — the same way they can't say which tools they pay for and never open. You can't price a dial you didn't know you were on.

That mapping is the first half of SMB Scale Up's AI Clarity Sprint: a fixed-fee, two-week engagement ($1,495 CAD) that inventories every AI touchpoint your business already pays for — the tools, the AI features embedded in software you own, and the shadow usage nobody approved — then pairs it with an own-vs-use analysis showing what's earning its keep, three ranked quick wins, and one first-project plan with a build/buy/don't verdict. You leave with documents you own, not a platform you rent.

If your AI bill went up and nobody can say why, start by finding out what you're actually running: https://smbscaleup.com/ai-clarity-sprint.html

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.