Some models produce internal reasoning (sometimes visible as “thinking”) before writing the final answer. They're much better at certain tasks and much worse value for others.
How they differ
- They generate extra reasoning tokens before the answer. These usually count as output tokens, so they cost money.
- They're slower: the first visible word may take noticeably longer.
- They're stronger on multi-step problems: maths, logic, tricky debugging, planning.
Use a reasoning model for
- Hard bugs where the cause isn't obvious
- Algorithm design and complex refactors
- Maths, data analysis with several steps
- Planning tasks with many constraints (schedules, migrations)
- Checking another model's answer for errors
Don't use one for
- Simple Q&A, summaries, rewriting, translation
- Classification and extraction at scale
- Chat where speed matters more than depth
For these, a standard model is faster and cheaper with similar quality.
Practical tips
- Give the problem, not the steps. These models plan well by themselves; over-specified prompts can hurt.
- Set a sensible max_tokens that leaves room for reasoning plus the answer.
- Compare on your task. Try a reasoning model such as
deepseek-reasoneragainstclaude-sonnet-4-6on five real examples and compare quality, time and cost in Usage Logs.
In DEVSAIO's AI Chat, models that think show their reasoning live as it happens, which is a nice way to get a feel for them.
Claude, GPT, Gemini, DeepSeek, image and video models through one OpenAI- and Anthropic-compatible endpoint. Prepaid, pay only for what you use.