Dial in how hard your agents think
New in v1.5.0: every agent has a Reasoning Effort setting that controls how much the model "thinks" before it answers. More thinking finds better answers on hard problems — but costs more tokens and time. Now you choose.
The three settings
- Off — no extended thinking. Fastest and cheapest. Great for lookups, formatting, short rewrites, and anything where the first answer is usually right.
- High — the balanced default. The model reasons through the problem without going overboard. A good fit for most real work.
- Max — the model thinks as hard as it can. Reach for it on genuinely hard problems: tricky math, multi-step logic, architecture decisions, thorny debugging.
Good to know
- The control only appears on models that actually support extended thinking (like GLM, Grok, and the OpenAI reasoning models). On models that don't, there's nothing to dial — so the UI never promises something the model can't deliver.
- Lower effort = fewer tokens = lower cost. If you're running an agent at scale, dropping from Max to High (or Off) is an easy way to save.
- It's a per-agent setting, so you can keep a deep-thinking research agent on Max while a quick formatting agent runs at Off.
A rule of thumb
Start at High. Drop to Off for the simple, high-volume stuff. Save Max for the problems that actually deserve it.
Found a sweet spot for a particular task? Share it in Tips & Tricks.