Best LLM API for Hermes Agent, measured

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

As of November 07, 2026, of the 14 LLM APIs built into Hermes Agent that we measure, OpenRouter answers fastest: 69 ms median time to first byte, averaged over 4 regions. By region, Fireworks AI is fastest from Europe (16 ms), US Central (29 ms) and Tokyo (19 ms), OpenRouter from São Paulo (62 ms). 4 of 14 hosts answered every probe over the 90 days since 2026-08-09; the lowest was OpenAI API (direct) at 99.1%.

Hermes Agent is the open-source, self-hosted AI agent from Nous Research that runs on your own machine or server and is reached from Telegram, Discord, Slack, WhatsApp, Signal, email or the CLI. Of the LLM providers it has built in, 14 are measured here at the very host Hermes Agent calls — directly, without a gateway, from 4 regions every five minutes. Any other OpenAI-compatible API can be added as the Custom endpoint option in hermes model.

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 Hermes Agent’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 Hermes Agent?

4 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).
• 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 Hermes Agent VPS runs?

For some APIs a lot. Upstage answers in 58 ms from Tokyo and 616 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 Hermes Agent calls exactly the host our probe measures. The same company on another address is another measurement, so these are left out:

• Nebius Token Factory: Hermes Agent calls api.tokenfactory.nebius.com, our probe measures api.studio.nebius.com (source)
• Tencent TokenHub: Hermes Agent 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: OpenClaw · n8n · VPS region for agents · data & licence