TL;DR: Reasoning effort is a dial, not a switch. Turn it up for planning and analysis, down for anything conversational, and set it per agent rather than globally.
This guide originally covered the adaptive thinking default that arrived in OpenClaw v2026.3.1, back when Claude 4.6 was current. Both have moved on. OpenClaw 2.0 shipped on September 1, 2026, the Claude model lineup is now on its fifth generation, and the way you express "think harder about this" has changed shape.
The underlying decision has not changed at all, which is why this page is worth updating rather than retiring.
When a model reasons before answering, it spends tokens working the problem internally before producing a reply. You pay for those tokens and you wait for them. In exchange you get better results on problems that genuinely require working through steps.
That trade is excellent on a planning task and terrible on "reply to this message." The whole skill is knowing which one you have.
Three things determine the outcome, and only one of them is the setting itself.
Model tier. Some models are built to feel quick, others to reason deeply. Your starting model still matters more than any flag. As of the current lineup, Fable 5.1 is built for long autonomous sessions, Opus 5 for complex reasoning, Sonnet 5 for daily work, and Haiku 4.5 for fast, simple tasks.
Effort level. Current Claude models accept an explicit effort setting rather than a simple on/off. The scale runs low, medium, high, xhigh, max on Fable 5.1, Fable 5, Opus 5, Sonnet 5, Opus 4.8, and Opus 4.7. Older models such as Opus 4.6 and Sonnet 4.6 support low through max without xhigh. This is the single biggest change since the original version of this guide: what used to be "adaptive or off" is now a five-position dial.
Agent role. A planner, researcher, or coding agent earns the extra thinking. A notification bot does not. This is the one that people get wrong, because it is tempting to set a good default once and apply it everywhere.
If you run OpenClaw through chat apps, remember that extra reasoning makes back-and-forth feel sluggish in a way that is very obvious in a live channel. The OpenClaw Multi-Platform Messaging Setup Guide is a useful companion, because response timing lands differently on each platform.
The original version of this guide showed a models.primary block with a thinking.mode field. That is not the current shape, so if you copied it, it is worth revisiting.
A current OpenClaw config declares models under a provider, and each model entry carries its own capability metadata, including whether it reasons:
{
"models": {
"providers": {
"openai": {
"api": "openai-responses",
"baseUrl": "https://api.openai.com/v1",
"models": [
{
"id": "gpt-5.5",
"name": "GPT-5.5",
"reasoning": true,
"input": ["text", "image"],
"contextWindow": 200000,
"maxTokens": 128000
}
]
}
}
},
"agents": {
"planner": { "model": "gpt-5.5" },
"messenger": { "model": "gpt-5.4" }
}
}The pattern to take away is the important part, because field names do move between releases: declare model capabilities once at the provider level, then assign models to agents by role. You express "this agent should think harder" by giving it a reasoning-capable model, not by scattering thinking flags across the file.
If you are not a developer, do not let the word "config" put you off. It is a settings file. You are editing preferences, not programming.
The asymmetry matters: the low-effort cases are usually the high-volume ones. Getting those right is where the cost savings live, not in shaving effort off the handful of hard problems where it actually pays.
Reasoning tokens are billed tokens, and on a busy agent they add up faster than people expect because the volume sits in the boring calls.
Two habits are worth building. First, measure before you tune. Look at which agents actually consume your tokens, then set effort per agent from evidence rather than intuition. Second, watch the tail, not the average. A planning agent that occasionally spends a long time on a hard problem is working correctly. A messenger agent doing the same is misconfigured.
Prompt caching is the other half of this, and it has moved recently. Anthropic cut cache reads to $0.25 per million tokens with Fable 5.1, a 75% reduction, which changes the economics of long agent sessions substantially more than effort tuning does.
How long the model is allowed to think before it answers. More thinking costs more and takes longer, and is worth it only when the problem actually has steps.
No. On simple, well-specified tasks it mostly adds latency and cost. It helps when a problem requires working through stages, and it does very little when the answer is a lookup or a reformatting job.
No, and the distinction is worth getting right. Fast mode is about responsiveness: in Claude Code it runs Claude Opus with faster output and does not swap you down to a smaller model. Effort is about how much reasoning the model does. One is a speed preference, the other an effort preference, and you can combine them.
No. That is the most common and most expensive mistake. Set it per agent, by role. Your planner and your notification bot should not have the same setting.
Assign a different model to that agent in your settings file, or override it for a single run from the command line. Neither requires writing code. If a field name in this guide does not match your install, keep the pattern and check your version's reference; OpenClaw releases move quickly.
low through max, with xhigh on the current generation.Have questions about running agents in production? Get in touch.
This article is a live example of the AI-enabled content workflow we build for clients.
This is a refresh of a March 2026 guide, rewritten in place rather than republished at a new URL.
| Stage | Who | What |
|---|---|---|
| Audit | Tom Hundley | Flagged the page from search data; it was pinned to a specific OpenClaw release and a superseded model generation |
| Research | Claude Opus 5 | Current model tiers and effort levels from vendor documentation; config shape checked against a live deployment |
| Drafting | Claude Opus 5 | Rewrote against verified sources |
| Fact-Check | Human + AI | Model names, effort levels, and pricing traced to primary sources |
| Editorial | Tom Hundley | Final review for accuracy, tone, and value |
What the refresh corrected. The original was written around OpenClaw v2026.3.1 and Claude 4.6, and framed thinking as adaptive-or-off; current models take a five-level effort setting. The example configuration used a models.primary block with a thinking.mode field that does not match the current schema, so it has been replaced with the shape a live install actually uses. A "Tomorrow's Preview" section promising a follow-up post has been removed, because this is a reference page rather than an episode.
Want to build this capability for your team? Let's talk about AI enablement →
Discover more content: