If you are building an AI product, the model you call and how long it thinks before answering are two of your biggest cost levers. They affect your margins, your latency and your quality. This guide turns Claude's published model guidance into a decision framework for founders who need to protect unit economics early.
The figures below come from Claude's own guide as summarized in the source material. Treat them as a starting point and re-measure on your own workload.
Two levers, not one
Every AI feature has two separate dials:
- The model: how big the brain is. This sets the price per token.
- The thinking level (effort): how long the model thinks before it answers. This, along with task difficulty, sets how many tokens a job uses.
Your cost per task is tokens multiplied by price per token. Founders who only compare price per token miss half the equation.
The four models as a product menu
The guide describes four models, which map neatly onto product jobs:
| Model | Role | Typical product use | Approx. price per million tokens written |
|---|---|---|---|
| Haiku 5.5 | Prep cook | Sorting and tagging at huge volume | about $0.50 |
| Sonnet 5.5 | Line cook | Clear, well-defined jobs | about $10 |
| Opus 5.5 | Head chef | Complex work; the default | about $20 |
| Fable 5.2 | Executive chef | Hardest, most ambitious technical work | about $40 |
That is roughly an eighty-fold spread from cheapest to most expensive. For a startup, that spread is the difference between a feature that scales profitably and one that does not.
Why the bigger model is sometimes cheaper
On a hard coding test, Opus 5.5 at max thinking scored about 45%. Fable 5.2 at high thinking matched that score for about half the cost per task. Pushing a smaller model to its maximum burns many tokens. So before you max out thinking on a model that is struggling, price the next model up.
What thinking levels buy you
Holding the model at Opus 5.5, the guide shows diminishing returns:
- Low: about a third of the cost of high, roughly eight points lower in score.
- Medium: about 70% of the cost of high, giving up only 2.5 points.
- High: the reference point for comparison.
- Extra high: 2.5 times the cost of high for 1.4 more points.
Medium is the default in the Claude apps, and it is a sensible default for your prototype too.
An escalation ladder for your roadmap
When outputs are not good enough, resist the urge to turn every dial up. Climb in this order:
- Let the model check its own work with a test or checklist.
- Raise thinking one level.
- Move up a model, and start it at a lower thinking level.
More thinking helps with planning, tool use and self-checking, but it cannot supply knowledge the model lacks. That gap needs a bigger model.
Two architecture patterns worth testing
Low first, retry the misses. Running everything on high scored 95% at 29 cents per task. Running low first and redoing only the misses scored about 97% at 17 cents, roughly 40% cheaper.
Lead plus helpers. An Opus 5.5 lead with Sonnet 5.5 helpers cost about a third less with about the same results. Give helpers jobs where mistakes are easy to spot, and keep judgment calls with the lead. This is a natural fit if you are building agents.
A founder's test plan
- Pick one real task from your product, not a benchmark.
- Run it on two models at a sensible thinking level.
- Record quality and cost per finished job, not per token.
- Set a default, and define what triggers a move up.
- Revisit when your volume or task mix changes.
This also gives you numbers for investor conversations: a clear cost-per-task and a plan for how it falls as you route work to the right model.
Go deeper
The original walkthrough with examples by role is in Which Claude model should you use?. If you are weighing these choices inside a larger product plan, the main website shows how Silverberry supports teams putting AI into production.