Which Claude model should you use?
aistrategy
Start on Opus 5.5 at medium thinking, raise thinking before switching models, and judge cost per finished job.
Claude's menu offers four models and several thinking levels, and the choice affects both quality and cost. The model is how big the brain is. The thinking level (called effort) is how long it thinks before answering. Claude's own guide says to start most work on Opus 5.5 at medium thinking, the default in the Claude apps.
The four models, as a kitchen
- Haiku 5.5 (prep cook): fastest and cheapest, for sorting and tagging at huge volume.
- Sonnet 5.5 (line cook): fast and capable, great at clear, well-defined jobs.
- Opus 5.5 (head chef): the daily driver for complex work, and the default.
- Fable 5.2 (executive chef): for the hardest, most ambitious technical work.
Per million tokens written, that's about fifty cents for Haiku, ten dollars for Sonnet, twenty for Opus and forty for Fable. But price per token isn't price per job. Your bill is tokens times price per token, and the thinking level and job difficulty drive the token count.
Thinking level and cost
On Opus 5.5, the guide's numbers show diminishing returns:
- Low: about a third of the cost of high, roughly eight points lower.
- Medium: about 70% of the cost of high, giving up only 2.5 points.
- Extra high: 2.5 times the cost of high for 1.4 more points.
Sometimes the bigger model is cheaper. On a hard coding test, Fable 5.2 at high thinking matched Opus 5.5 at max (about 45%) for roughly half the cost per task.
When the answer isn't good enough
- Let it check its own work with a test or checklist.
- Raise the thinking one level.
- Move up a model, and start it at a lower thinking level.
More thinking helps with planning, tools and self-checking, but it can't add knowledge the model lacks. Two other patterns from the guide: running low first and redoing only the misses scored about 97% at 17 cents per task versus 95% at 29 cents for everything on high. An Opus 5.5 lead with Sonnet 5.5 helpers cost about a third less with similar results.
Before you choose an AI tool
Ask which model and thinking level it uses by default, whether it moves up only when the job needs it, and whether you can see what each task costs. Then test one real task on two models and compare result and cost. For help turning this into AI strategy for your organization, Silverberry's AI consulting services are one conversation away.