By Scott Covert · November 14, 2026
- The model decides how much Claude knows. The thinking level decides how hard it tries.
- Turning thinking up never teaches a model what a bigger model knows.
- The cheapest good result usually comes from a smart lead and cheap helpers, not one big model doing everything.
Most people set both dials once and never touch them
The usual habit is one model, one setting, for everything.
That means paying Fable prices to sort email, or asking a small model for judgment it doesn't have.
Anthropic now sells four current models. From smallest to biggest: Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.2.
Each step up roughly doubles the price per token from Sonnet on. Haiku sits far below the rest.
On top of that, each model can think at five levels: low, medium, high, extra high and max.
More thinking means Claude reads more, tests more and makes more decisions on its own. It also means more time and more output tokens, which you pay for.
Raise the thinking level one notch before you jump to a bigger model.
It's the smaller step. When an answer misses, ask which kind of miss it was: Claude didn't know enough (go up a model) or it didn't try hard enough (go up a level).
What each model is for
Haiku is the fastest and cheapest. Give it summarizing, sorting, tagging and routing: simple, repetitive work that's easy to check.
Ben AI, reporting Anthropic's own tests, says Haiku 5.5 scored 85% on a hard science quiz against 91% for Opus, at about a twentieth of the cost per correct answer. The same report says Haiku fell off by almost a fifth on hard search tasks that need judgment calls.
Sonnet costs half what Opus does per token. It's the pick for scoped tasks where "done" fits in one sentence, and for the read-only helpers that go and look things up.
Opus 5.5 on medium is the model Anthropic recommends as the daily driver. It's also a strong lead: it plans the job and hands scoped pieces down.
Fable 5.2 costs twice what Opus does. Use it for ambitious, open-ended work where planning, judgment and accuracy carry the result, and for leading big teams of helpers.
An Opus lead with up to 25 Sonnet helpers cost about a third less than Opus doing everything alone, and scored about the same.
That's from Anthropic's own testing, as Ben AI reports it. The work didn't get worse when the legwork moved down a tier.
How the hand-off works, step by step
Every example in the demo above follows the same shape.
- The lead reads the job and splits it into pieces that each have a clear finish line.
- Helpers run at the same time, each on the cheapest model that can do its piece.
- Anything unsure goes up, not everything. Only the hard cases reach the expensive model.
- Results come back to the lead, which checks them and writes the finished thing.
In the bug-fix example, the first fix fails its tests. The lead raises its own thinking level instead of handing the job to Fable. That's the one-notch rule in action.
Making it almost automatic
You don't have to choose by hand every time. In Claude Code, each helper can carry its own model and thinking level in a small settings file, so the lead picks the right one by itself.
--- name: code-searcher description: Read-only search through the codebase. Finds files and lines; changes nothing. model: sonnet ---
Skills can name a model the same way, so one task runs on Sonnet while the rest of the chat stays on Opus.
Try the five questions below. They're the same rules a lead model can follow on its own.
Which cell should this task start in?
One more trap: switching models halfway through a chat. A saved copy of the conversation belongs to one model only, so the new model re-reads everything at full price, and moving down a tier drops the earlier reasoning.
The fix is a short hand-off note: start a fresh chat on the new model with a summary of where things stand.
- The token counts in the demo are illustrative, not measured.
- Max thinking adds a lot of cost for small gains and can overthink.
- No setting rescues a vague request. A better brief beats a bigger model.
- Most people never learned to delegate, to people or machines. Handing a piece of work to a cheaper helper feels like losing control.
- Picking one model and leaving it there is the same habit as doing it all yourself because explaining it takes longer.
- The brief is the bottleneck. Someone who can't say what "done" means can't get it from Haiku or from Fable.
My read
The models got good enough that the skill now is managing them: who leads, who does the legwork, and when to turn a dial.
That's an old skill. Anyone who has run a small team already knows most of it.
The matrix as a table
| Model | Price per million tokens (in / out) | Best for | Usual setting |
|---|---|---|---|
| Fable 5.2 | $8 / $40 | Open-ended work, planning, leading many helpers | High |
| Opus 5.5 | $4 / $20 | Daily driver, judgment, writing, leading helpers | Medium |
| Sonnet 5.5 | $2 / $10 | Scoped tasks, read-only helpers | Low to medium |
| Haiku 5.5 | $0.10 / $0.50 (prompts up to 100K tokens) | Sorting, summarizing, tagging, routing | Low |
Questions
- Does a higher thinking level make a smaller model as smart as a bigger one?
- No. Sonnet on max still doesn't know what Opus knows. Thinking changes effort, not knowledge.
- Should I raise the thinking level or switch models first?
- Raise the level one notch first. It's cheaper, and switching models mid-chat means paying for the whole conversation to be re-read.
- Which model should run helper agents?
- Usually Sonnet, and Haiku for simple work that's easy to check. Keep the bigger model as the lead.
Sources
- Anthropic, Claude pricing (list prices, read November 11, 2026).
- Ben AI, Copy this Claude Mod, Never Overpay Claude Again (YouTube, November 14, 2026), reporting Anthropic's guidance and test results on models, thinking levels and helper agents.