Best LLM API for OpenClaw, measured

Updated November 07, 2026 · built-in providers checked against OpenClaw’s documentation on 2026-11-07

As of November 07, 2026, of the 22 LLM APIs built into OpenClaw that we measure, OpenRouter answers fastest: 69 ms median time to first byte, averaged over 4 regions. By region, Fireworks is fastest from Europe (16 ms), US Central (29 ms) and Tokyo (19 ms), OpenRouter from São Paulo (62 ms). 6 of 22 hosts answered every probe over the 90 days since 2026-08-09; the lowest was OpenAI at 99.1%.

OpenClaw is an open-source AI assistant that runs on your own hardware and answers in the chat apps you already use — WhatsApp, Telegram, Slack, Discord and more. Of the LLM providers it has built in, 22 are measured here at the very host OpenClaw calls — directly, without a gateway, from 4 regions every five minutes. Any other OpenAI-compatible API can be added as a custom provider under models.providers with its own base URL.

How fast and reliable is each built-in provider?

Median time to first byte over the last 24 hours from each probe region, the share of host probes answered over the 90 days since 2026-08-09, and outages we could attribute to the provider in the last 30 days — of the host, and of real requests to one model where we hold an API key. Bold marks the provider’s fastest region; an outage count links to its latest day.

How long until the first token?

Time to first byte is the network floor under every call. The wait for the model’s first token is longer, and we time it only where we hold an API key and send a real request — for these built-in providers, median time to first token over the last 24 hours:

More: time to first token by provider.

Which built-in provider is cheapest for the same model?

Across different models a price comparison means nothing, so this table holds the model fixed: open-weight models that two or more of OpenClaw’s built-in providers serve, with what each charges per million input tokens and which of them answered fastest. Prices are published by the providers and read through OpenRouter’s public catalogue on 2026-11-07; they are not measured by us.

Is there a free LLM API for OpenClaw?

5 of the built-in providers can be used without paying, within limits (checked 2026-11-07 on their own pages):

• Google Gemini — free input and output tokens on a limited set of Gemini models; rate-limited; content sent on the free tier is used to improve Google's products; apps for users in the EEA, Switzerland or the UK must use paid services (terms) (terms).
• OpenRouter — model variants whose ID ends in :free; 20 requests a minute; 50 a day, or 1,000 a day once you have bought at least 10 credits (terms).
• Groq — a Free plan; a payment method is needed only to move up to the Developer tier; per-model rate limits, published on Groq's rate-limits page (terms).
• Z.AI (GLM) — GLM-4.7-Flash, GLM-4.5-Flash and GLM-4.6V-Flash: input and output both free; rate limits not published in the docs; the other GLM models are paid (terms).
• NovitaAI — one model priced at $0 for input and output: Apodex 1.1 Mini; rate limits and how long it stays free are not stated (terms).

One-time credits for new accounts and every other free option we found: free LLM APIs, measured.

Does it matter where the OpenClaw VPS runs?

For some APIs a lot. Qwen Cloud answers in 167 ms from Tokyo and 706 ms from São Paulo, and every model call waits at least that long before the model starts working. The bold figure in each row of the first table is that provider’s fastest region; put the agent there. Which VPS region is closest to the LLM APIs compares all four.

Why are some built-in providers missing?

A figure is shown only when OpenClaw calls exactly the host our probe measures. The same company on another address is another measurement, so these are left out:

• Cohere: OpenClaw calls api.cohere.ai, our probe measures api.cohere.com (source)
• Meta: OpenClaw calls api.meta.ai, our probe measures api.llama.com (source)
• Tencent TokenHub: OpenClaw calls tokenhub.tencentmaas.com, our probe measures api.hunyuan.cloud.tencent.com (source)

How are these figures measured?

Each probe opens a connection to the provider’s API host and times the first byte of the answer, without an API key: a 401 counts as reachable, because the host answered. Probes are rented VPSes — Hetzner in Nuremberg, Vultr in Chicago, Vultr in Tokyo, Vultr in São Paulo. Host uptime is the share of probes that got a valid answer. An outage is a failure seen from two or more regions at once that no other provider shared in the same minutes; a failure from one region alone is not called one. Where we hold an API key we also send a real one-token request to one model, and its failures are counted separately from the host’s. Time to first byte says nothing about tokens per second once the model runs, or about model quality. Full rules: methodology.

Other agents: Hermes Agent · n8n · VPS region for agents · data & licence