Hermes Agent pricing: what running it actually costs

Hermes Agent costs $0 to install: the software is MIT-licensed open source and the only unavoidable bill is model tokens. Everything else is optional and priced separately: a Nous Portal subscription at $0, $20, $100 or $200 per month, Hermes Cloud hosting from $0.56 per day, and Tool Gateway usage, a paid-subscription feature that publishes no per-use rate.

This page is about Hermes Agent, the open-source agent from Nous Research. A live check of the repository on September 21, 2026 returned archived: false, an MIT license and a push that same day; the newest published release is Hermes Agent v0.21.3, tagged v2026.9.14 on September 14, 2026. It is not Hermes 3 or Hermes 4, the Nous open-weight model family, and not the luxury-goods brand of the same name.

Four separate bills, only one of which is compulsory

Hermes Cloud bills the instance daily in arrears, and its own page states that inference and tool usage are billed separately on top of the instance price. A hosting quote is therefore never the whole bill. Local models, per the same FAQ, are completely free to run once you own the hardware. All figures checked September 21, 2026. The same four-bill breakdown applied to a different agent is in OpenClaw pricing.

Nous Portal plans: you are buying credits, not a token discount

The credits are the product. A $20 subscription buys $22 of model spend, so its advantage over pay-per-token is the bonus plus the bundled tools and limits — not a cheaper rate on every model. On the public catalog read on September 21, 2026, Claude entries carried the rates Anthropic publishes — Opus 5 at $5.00 in / $25.00 out, Fable 5.1 at $10.00 / $50.00 — with no reduction shown. Whether anything else is discounted is not something that page lets you check: it prints one rate per entry and no reference price beside it. What it does show is that one model can appear several times at different rates, because regional and batch variants are listed as separate entries — Kimi K3 at $1.70 / $8.50, its :US entry at $3.63 / $18.15, its batch entry at $3.00 / $15.00. Budget against the exact entry id you will call, and date-stamp anything you quote.

Three caveats before you budget. The free tier is real but restricted to the entries whose model id carries a :free suffix — seven of them on the September 21, 2026 read — and that list is rebuilt from the live catalog, so “use the free one” advice dates quickly. Portal's own surfaces quote two different round figures for the catalog size — a smaller one on every paid plan card, a larger one in the section above the catalog itself — while the published list held 360 entries on that read, so treat any of those figures as marketing rather than as a budget input. No request-rate numbers appear on the plan cards either; they say only “standard” and “high” rate limits, so treat concurrency as unverified until you see it in your own account. The plan table lives on the Portal homepage and on manage-subscription, where a plain HTTP client gets a Vercel security checkpoint instead of the page — read it in a browser.

Where the tokens actually go

A rate card prices one call. Hermes bills a workload, and three documented multipliers sit between the two.

Auxiliary slots. The dashboard exposes 11 auxiliary task slots — title generation, vision, compression, approval, web extract, skills hub, MCP routing and others — and every one defaults to auto, meaning Hermes tries your main model for that job too. The documentation's own example of waste is compression: a fast chat model does that job at one fiftieth of the cost of a reasoning model. One user turn can therefore be several billed calls.

Delegation. The delegation documentation states that a parallel batch of subagents typically burns the large majority of a run's total tokens, which is why the worker model is configurable separately as delegation.model. Its own cost warning notes that max_spawn_depth: 3 with max_concurrent_children: 3 can reach 27 concurrent leaf agents; the default depth is 1.

Cache continuity. The same model configuration page warns that prompt caches are keyed to the model serving the request, so any mid-conversation model change — an explicit /model switch, an automatic fallback, or a credential-pool rotation onto a different account — makes the next message re-read the whole conversation at full input price instead of the cached rate, which it puts at roughly 75 to 90 percent discounted. A default guard, model.switch_context_confirm_tokens: 100000, exists precisely because that one-time re-read can dwarf the per-token difference between the two models.

Scheduled work is where this bites hardest, and not in the direction you would hope: the cron documentation says a job with no model of its own runs on whatever hermes model or /model is set to at the moment it fires, so changing your chat model moves the whole cron fleet with it on its next run. Pin the job (hermes cron create/edit --pin, or an explicit --model / --provider) or set a fleet default with cron.model when an unattended job must not follow a switch. Script-only jobs — the doc's no_agent mode — never touch the inference layer at all. Rather than estimate any of this, measure it: the FAQ documents /usage, /compress and hermes prompt-size, and the web dashboard documents an analytics view with total tokens, cache-hit percentage and a per-model breakdown.

Choosing the model for each job

No benchmark here supports a quality ranking, so this is a table of what decides each choice, not a leaderboard.

Two official positions are worth knowing. Nous ships a low-cost default: the remote model manifest, stamped September 20, 2026 and read the next day, marks z-ai/glm-5.2 as default, and the catalog documentation says the default is deliberately a capable low-cost model, never the priciest flagship. And Nous advises against its own weights inside the agent: the Portal integration page says Hermes 4 is not recommended for use inside Hermes Agent because it is tuned for chat and reasoning rather than the rapid-fire tool-calling loop. That page's shortlist — Claude Sonnet 4.6 as the general-purpose agentic pick, GPT-5.5 Pro, a long-context Gemini 3 Pro preview, DeepSeek V4 Pro as the cost-effective coder — lags both the manifest and the live catalog, so treat it as a starting point and confirm the model id resolves.

Observed usage is not a ranking either. OpenRouter's public app page for Hermes Agent, read September 21, 2026, showed 30-day token volume led by MiniMax M3, an unlabelled entry called Ox Alpha, Ling 3.0 Flash Fin, GPT-4o-mini and two DeepSeek V4 Flash builds — cheap, fast models rather than flagships. That records what people ran through one gateway, not what performed best on your task.

“Hermes models” is two different questions

One is which model to point the agent at, answered above. The other is the Hermes 4 open-weight family — Hermes-4-405B, -70B and -14B on Hugging Face, with Hermes 3 distributed through Ollama. Those are model weights from the same company, not the agent, and Nous says not to run Hermes 4 inside it. One claim here is unverified: the Portal integration doc says Hermes 4 models are available on the Portal at heavily discounted rates, but a search for “hermes” and “nous” across the live public catalog returned no matching rows on September 21, 2026. Either they are visible only when signed in or the doc is stale. Do not budget against a Hermes 4 Portal price you cannot see in your own account.

Best and cheapest API for Hermes Agent

These are different claims and should stay apart. Cheapest listed price is a rate-card fact; lowest cost to finish the task depends on retries, context length and cache hits.

Hermes documents that any service with an OpenAI-compatible API works, and its provider reference even carries recipes for competing gateways, so the custom-endpoint path is a first-class route rather than a workaround. Where Kunavo fits is narrower than “cheapest”: its catalog covers Claude, Gemini and GPT-class models plus media models, and the open-weight models that dominate observed Hermes usage are not on it. If your setup already runs on those, Kunavo is not the cheaper route. It is the route for keeping Claude, Gemini or GPT main turns at a listed rate under one key.

The same model, the same tokens, two price lists

USD per million standard uncached input and output tokens. Anthropic's direct rates come from its pricing page checked September 21, 2026; the Kunavo column reads the live catalog.

At today's catalog rates, Claude Opus 5 and Claude Sonnet 5 carry the same per-token price on Kunavo for both input and output, even though they do not carry the same price direct from Anthropic. For a new Hermes main slot, that makes the choice between them a capability and response-behavior question rather than a budget one.

Now the arithmetic. Assume a month of 20 million uncached input tokens and 2 million output tokens, counted across main turns, the auxiliary slots and delegation workers together — that combined count is the point, because a per-turn estimate misses two of the three.

This is token arithmetic at catalog rates, not a measured Hermes task and not a bill ceiling. It excludes cache writes — Claude cache writes bill at 1.25 times the input rate here, explained in the caching doc — and excludes paid tools, hosting and tax. It also assumes the cheaper model finishes its assigned work without extra attempts. That assumption, not the rate, is what decides whether the cheaper row is actually cheaper; cost optimization covers how to test it.

Kunavo's catalog amount is a billing floor rather than a cap: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup. The minimum is a $10 prepaid top-up with no subscription, which is the cash needed to open a funded account, not the cost of a task. See billing.

Wiring a third-party endpoint into Hermes

This configuration comes from Hermes' own provider reference, reviewed as source documents on September 21, 2026. No Hermes run against Kunavo has been performed, there is no Kunavo setup guide for Hermes, and nothing below should be read as a tested integration. Keep your working route available while you try it.

# ~/.hermes/config.yaml
providers:
  kunavo:
    api: https://api.kunavo.com/v1
    key_env: KUNAVO_API_KEY
    transport: chat_completions   # set it; do not rely on URL auto-detection
    models:
      claude-sonnet-5:
        prompt_caching: true

model:
  default: claude-sonnet-5
  provider: custom:kunavo

KUNAVO_API_KEY=your-key

Five things a custom endpoint does not get for free, each of which has a documented cost:

1. Set transport explicitly. Leaving it blank falls back to URL-based auto-detection, which the provider reference describes as a fallback for a blank field and illustrates with exactly one rule — an /anthropic path resolving to anthropic_messages. The rest of that heuristic is not enumerated on any page read here, which is the reason to name the transport rather than to assume nothing will fire. Hermes' own regression test for that path records the failure mode it guards against: an openai-api entry pointed at a regional OpenAI host silently resolved to chat_completions rather than the transport the provider itself declared, and every tool-calling turn returned a 400. Kunavo serves /v1/chat/completions, /v1/messages and /v1/responses, so all three documented transports have a matching endpoint on paper — a documentation-level match only, not a tested one. Name the transport you want instead of letting detection choose it.
2. Opt each model into prompt caching. For a gateway that resolves a bare model alias, cache markers must be declared per model; Hermes states it will not infer cache support from a provider's name, host or model family. Missing this silently forfeits the cached-input rate, which is the largest lever in a long session. See prompt caching.
3. Set context_length when the default is wrong. Custom OpenAI-compatible endpoints receive no automatic catalog-sized output cap — the server's own defaults apply, and they can be lower than the model maximum. The context window resolves through a chain that ends at models.dev and then a 128K fallback.
4. Know what extra_headers covers. The model configuration page applies it to OpenAI-compatible routes and to anthropic_messages routes — main client, /model switches, rebuilds and auxiliary clients alike — and names bedrock_converse as the one mode that does not use it. It says nothing either way about codex_responses, so treat that combination as untested. On the Anthropic wire the provider doc's example points api at a path without stating the suffix Hermes appends, so check the recorded request path on the first call — the base-URL doc covers the origin-versus-/v1 trap behind most 404s there.
5. Use hermes model to add a provider. Mid-session /model switches between endpoints you already configured and cannot add a new one — and that switch discards the cache.

For a first trial, run one bounded task with a real tool call, then read /usage and the dashboard breakdown before moving scheduled work across. Create a Kunavo account when you are ready to test that route with a funded key. Still choosing the agent rather than its provider? Hermes vs OpenClaw compares the two workflows.

FAQ

Nous Research and Anthropic pages checked September 21, 2026; the repository state and latest release were verified against the GitHub API the same day. Kunavo rates and every total above are computed from the live model catalog and are illustrative estimates. Hermes configuration is reported from source documents, not from a runtime test.