cd /news/large-language-models/a-more-expensive-ai-isnt-always-more… · home › topics › large-language-models › article
[ARTICLE · art-172895] src=somethingbig.ai ↗ pub= topic=large-language-models verified=true sentiment=· neutral

A more expensive AI isn’t always more expensive

A newsletter author writing as Matt argues that cheaper-per-token AI models such as Sonnet can cost more per finished task than pricier models like Opus, because the cheaper model needs more rounds of correction and more tokens to complete the same job. Matt cites a marketing-page example in which Sonnet costs about half as much per token but needs three rounds of feedback, totaling roughly one and a half times Opus's cost, and recommends running one real task on both models while recording cost, time and number of corrections. Matt also urges AI labs to report cost per task instead of cost per token.

recommendation

Run one real task on both the cheap and expensive model, recording cost, elapsed time and number of corrections, before switching to the cheaper model.

more in playbook: Controlling LLM cost
by read3 min views6 publishedNov 19, 2026
A more expensive AI isn’t always more expensive
Image: Somethingbig (auto-discovered)

A lot of you have had this conversation with your boss: “Use Sonnet instead of Opus or Fable. The results won’t be as good, but we can’t afford the expensive one.”

It sounds reasonable. Fable costs more than Opus, and Opus costs more than Sonnet. But the sticker price doesn’t tell you the whole story. A lot of the time, the cheaper model ends up costing you more.

Here’s how. AI models charge by the token, which is roughly one word they read or write. The sticker price is the price per token. But you’re not buying tokens. You’re buying a finished job.

Think of it like hiring. Say you have a really hard physics problem. You could pay a college student $20 an hour, or you could pay Einstein $1,000 an hour. The student spends 100 hours on it and may still get it wrong. That’s $2,000. Einstein solves it in an hour. That’s $1,000, and you know it’s right.

It’s the same reason a senior employee can be cheaper than a junior one. And it’s true of AI models too. A smarter model takes fewer steps and makes fewer mistakes, so it uses fewer tokens to finish the job. You pay more for each token, but you buy far fewer of them.

Here are three places this shows up.

Do-overs. Say I want a new marketing page based on my existing ones. Opus might nail it on the first try. Sonnet might need three rounds of feedback. Sonnet costs about half as much per token, so three rounds comes to about one and a half times what Opus cost me. And that’s before you count the time I spent writing the feedback.

Effort settings. Most models let you choose how hard they think. A lot of people pick the cheaper model, then turn the effort all the way up to make up the difference. In my experience, you’re usually better off with Opus on low or medium than Sonnet on high or extra high. The work is way better, and for most tasks it ends up cheaper.

Speed. Sonnet writes faster. But Opus needs to write less to finish the job, and it has fewer mistakes to go back and fix. So the “slower” model often gets you there sooner.

To be fair, this doesn’t always go one way. On simple jobs, the cheaper model often does win. That’s exactly why you should test it instead of guessing.

So before you switch to the cheaper model, take one real task from your week and run it on both. For each one, write down three things:

1. What it cost (or how much of your plan it used).

2. How long it took.

3. How many times you had to correct it.

Then go back to your boss with the numbers. There’s a good chance you get to say: “The cheaper one did worse work, took longer, and cost us more.”

I hope the AI labs start talking about price this way, as cost per task instead of cost per token. Maybe that happens this year, maybe not. Either way, you should already be thinking about it like this.

Hit reply and tell me what you find. I read every one.

— Matt

What should I write about next? Vote or suggest a topic →

── more in #large-language-models 4 stories · sorted by recency
── more on @sonnet 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/a-more-expensive-ai-…] indexed:0 read:3min 2026-11-19 · —