You Probably Don't Need the Newest AI Model
Anthropic launched its newest top model this week and told users to start with a different one for everyday work. Here's how to match each job to the right model.

On Monday, Anthropic released Claude Fable 5.2, its newest top model. In the same announcement, the company told customers to start with a different one for most work. "For most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5, our strongest daily driver," the company wrote. Fable, it said, is "best for the most ambitious, open-ended technical work."
The same day, Meta went after the other end of the market. Its new small model, Muse Ember, costs 10 cents per million input tokens and 50 cents per million output, and Meta says it matches Claude Haiku 5.5, Anthropic's small model, on Artificial Analysis's intelligence index at about a third of the cost per task. (That's Meta's own chart, so read it the way you'd read any company grading itself.)
Put the two side by side, and the spread is wide: Fable's output costs $40 per million tokens, Ember's costs 50 cents. That's 80 times the price, and most of what a small business asks AI to do — answer the customer email, pull the totals off an invoice, summarize the meeting — sits a lot closer to the bottom of that range than the top.
So before anybody upgrades anything, here's how I'd sort the work.
The steps
- List the jobs, then price the mistakes. Write down what you and your staff use AI for in a normal week. Next to each job, write what happens when the answer is wrong: who catches it, and what it costs if nobody does. A clunky subject line costs nothing. A wrong number on a quote can cost you the job.
- Start everything on the everyday model. On a Claude plan, that's Opus 5.5, which is Anthropic's own pick for most work. On any other app, it's the model the company recommends for daily use, not the newest name on the menu.
- Push the busywork down. The jobs that run all day and cost little when they're wrong (sorting email, tagging receipts, pulling fields off a PDF) belong on the small models. Anthropic points people to Sonnet 5.5 and Haiku 5.5 for lighter work, and Meta is pitching Ember for repetitive agent work like routing and pulling data out of documents. If you pay per use, try the small step on two vendors before you commit to one.
- Turn up the effort before you trade up. Medium effort is now the default in Claude's apps and in Claude Code. When the everyday model gets something wrong, switch it to high and run it again before you reach for a bigger model.
- Give the top model a tryout, not your whole workload. This week Anthropic is giving subscribers on eligible paid plans extra Fable usage, and on Pro plans and standard Team seats Fable can take up to half your weekly limit. Use it to rerun one hard job from last month (the contract comparison, the messy reconciliation) on both models, and keep whichever answer is right, not whichever sounds smarter. If Fable doesn't win, you just saved yourself the upgrade.
- If you pay per use, check the first bill. Fable costs $8 per million input tokens and $40 per million output. Its fast mode doubles that, to $16 and $80, for answers up to 2.5 times faster. Look at what a week actually costs before you make it the default for anything.
How I run it
For what it's worth, this is how I run Palamo's own agents: Opus plans the work and hands it out, Sonnet is the workhorse, Haiku is the mule, and Fable is my master reviewer and diagnostician. I leave Fable on medium effort.
Anthropic's own launch post says to start with Opus 5.5. For most small-business work, I'd stay there until a job proves it needs more.
Your turn
Tell me what you use AI for in a normal week, and I'll tell you which of those jobs I'd give to the top model.
New files every week. Free.
Subscribe