Try Free

GLM Effort Levels: Non-Thinking vs High vs Max

GLM effort levels are the flagship GLM model's per-request control over how much computation it spends on a task. There are three — Non-Thinking, High and Max — and choosing between them is the single most consequential knob you have, because on the published curve the difference between the cheapest and most expensive setting is roughly eleven points of accuracy and 2.4× the output tokens.

A GLM model does not have one speed. Treating effort level as a per-task decision rather than a global setting is what makes the feature worth having: routine work runs cheaply, hard work gets the compute it needs, and nothing pays for depth it does not use. The figures below are as published by the model authors.

The Three GLM Effort Levels

The published curve, averaged over Terminal-Bench 2.1, DeepSWE and SWE-Atlas QnA on Claude Code 2.1.167, runs roughly as follows. The token figures are average output tokens per task, which is the cost that actually varies.

Non-Thinking

~63%
~35K avg tokens

Fast edits, boilerplate, routine refactors where latency matters more than depth.

High

Most used
~72%
~43K avg tokens

The default for most agentic coding: near-Max quality at roughly half the token spend.

Max

~74%
~83K avg tokens

Hard, long-horizon problems — allocate extra compute when a task genuinely warrants it.

Averaged over Terminal-Bench 2.1, DeepSWE and SWE-Atlas QnA, evaluated on Claude Code 2.1.167. Reported by the model authors. At comparable token budgets GLM-5.2 sits between Claude Opus 4.7 and Opus 4.8.

Reading the GLM Effort Curve

The two steps are not equivalent. Moving from Non-Thinking to High buys about nine points of accuracy for about 23% more output tokens — an unambiguously good trade for almost any task you care about. Moving from High to Max buys about two more points for roughly 93% more tokens, which is a good trade only when those two points decide whether the task succeeds at all.

That asymmetry is why High is the practical default. It also explains why blanket-enabling Max is a common and expensive mistake: on a queue of routine tasks it nearly doubles token spend to gain almost nothing, and on genuinely hard long-horizon problems it is the setting that makes the difference. At comparable token budgets, the model authors position GLM-5.2 between Claude Opus 4.7 and Opus 4.8. It is also worth noting what the curve does not say. These are averages over three evaluations, so the crossover point on your own workload could sit anywhere — a task distribution dominated by short edits will see even less benefit from Max than the average suggests, while one dominated by multi-hour debugging may see considerably more. The shape of the curve transfers; the exact numbers do not.

Choosing a GLM Effort Level in Practice

Route by task class rather than by preference. A workable default policy looks like this:

  • Non-Thinking for fast edits, boilerplate, formatting, routine refactors and anything where latency matters more than depth.
  • High as the default for agentic coding — near-Max quality at roughly half the token spend of Max.
  • Max for hard, long-horizon problems: multi-hour runs, deep debugging, architecture work and anything where a failed attempt costs more than the extra tokens.
  • Escalate rather than start high: run High first, and retry at Max only on failure. That keeps the expensive setting on the small fraction of tasks that need it.
  • Measure on your own workload. The published curve averages three evaluations; your task distribution will have its own shape and its own crossover point.

All figures on this page are results as published by the model authors. glmmodel.com reports them; it does not run these evaluations.

GLM Effort Levels FAQ

What are GLM effort levels?
Non-Thinking, High and Max are per-request settings on the flagship GLM model that control how much computation it spends on a task. On the published agentic coding curve that is roughly 63% at about 35K output tokens, 72% at about 43K and 74% at about 83K.
Which GLM effort level should I use?
High for most agentic coding — it is near-Max quality at roughly half the token spend. Non-Thinking for fast, routine work where latency dominates. Max only for genuinely hard, long-horizon problems where the extra two points of accuracy decide the outcome.
Is GLM Max effort worth the token cost?
Only selectively. Max costs about 93% more output tokens than High for roughly two more points on the published curve. That is excellent value when a failure is expensive and poor value on a queue of routine tasks, which is why escalating on failure beats defaulting to Max.
Do older GLM models have effort levels?
No. Explicit effort levels were introduced with GLM-5.2. GLM-5.1 and earlier releases spend a fixed amount of computation per task, so anything built against them cannot dial the trade-off per request.
How does the effort curve compare to closed models?
The model authors position GLM-5.2 between Claude Opus 4.7 and Opus 4.8 at comparable token budgets. The measurements were averaged over Terminal-Bench 2.1, DeepSWE and SWE-Atlas QnA and evaluated on Claude Code 2.1.167.

Keep reading

Try a Live AI Model Free — No Account Needed

Read the GLM model benchmarks, then put a real model to work. The playground above is free and needs nothing from you; if you want a fuller AI toolkit, our partner's free tier starts here.

Try FreeBack to the playground →