Skip to content
NLEN
Illustration: a pair of scissors cuts a long strip of code coming out of a terminal window in half, next to a gear with a slider set low.

OpenClaw and the 8,192-token limit: why replies stop halfway

By Ivo Donker · — text: Grok Bot (AI), checked against the sources by Claude (AI)

Why does an OpenClaw reply sometimes stop in the middle of a tool call? Until recently, OpenClaw filled in a maximum of 8,192 output tokens when a self-configured model had no maxTokens and the model was not in the bundled catalog. Long replies were cut off without a clear error. The npm package of version 2026.11.3 uses a new default of up to 32,768 tokens for most configurations. An explicit maxTokens per model remains the safest setting.

We read the OpenClaw documentation, an open bug report and the merged code change on GitHub, and the Harness Leaderboard changelog on Hugging Face. The topic was also in the trending briefing of November 13.

What went wrong

In bug report #196544 (November 10, still open), a user describes how a model without maxTokens in models.providers got a hard cap of 8,192 output tokens, while the provider allows 128,000. There was no warning when the configuration loaded. A long agent turn stopped at exactly 8,192 tokens, in the middle of a tool call, with only the message "Provider completed stream with an incomplete tool call". The reporter sees it only for model IDs that are not in the bundled catalog; a known model takes its limit from the catalog.

What it cost in a benchmark

The Harness Leaderboard changelog by Joël Niklaus reports that OpenClaw asked for at most 8,192 tokens per reply there, because their configuration set no output limit. On Flint and DeepSeek-V4.1-Flash that cut replies off and often ended the run. After a re-run with OpenClaw 2026.11.1, with each model's context window, longest reply and image input passed in, scores on DeepSeek-V4.1-Flash rose by 16 to 21 points. On GLM-5.3-Flash they fell by about 3 points. That gain comes from several settings together and a newer version; the output limit is one of them.

What OpenClaw changed

On November 6, pull request #193899 was merged. It replaces the filled-in 8,192 with a default of up to 32,768 tokens, smaller for a small context window. The custom provider documentation now describes it that way. We found the new default in the npm package of 2026.11.3 (November 10), although the release notes do not mention it. According to the PR itself, saved values of 4096 or 8192 from an older install still apply until you change them, and a few code paths still use their old fallback. A strict provider may reject 32,768 with a 400 error; in that case, set that model's real limit.

What this means

Anyone running OpenClaw with their own or local models should go through the model configuration and set a maxTokens on each model that matches what the provider allows. The documentation recommends this too. Also look for old values of 4096 or 8192 saved during setup. We covered earlier OpenClaw updates in our piece on the 2026.11.2 patch. Why the balance between context and output matters is explained in the race for context.

What we did not check

We did not reproduce the limit ourselves. That 2026.11.3 contains the new default is based on the code in the npm package, not on a release note. The benchmark figures come from the Harness Leaderboard itself.

Checked on November 13, 2026.

This article was written with AI (Grok Bot) and checked against the sources below.

Sources