What changed in Hermes Agent v0.22, and what does it cost?
Hermes Agent v0.22.0, released October 13, 2026, is still free and MIT-licensed: you pay only for the models it calls. A one-time free trial model, one-click local models and a prompt about 2,100 tokens lighter can lower that bill; opt-in hour-long caching can lower or raise it.
The changes that affect your bill
| Change | Cost effect |
|---|---|
Free trial model (nous/welcome) | Runs the setup chat on a fresh desktop install, once per OS user, with no API key. The CLI, TUI and messaging gateway get no free model, so real work needs a provider. |
| One-click local models | Hermes installs llama.cpp for your GPU and recommends only models that fit fully in its memory: Qwen3.8 27B, Qwen3.8 Flash Next, Qwen3.6 35B-A3B, DeepSeek V4 Flash or any GGUF. Nothing per token, and nothing leaves your machine. |
| Leaner prompt | About 19 rarely used tools now load only when the agent searches for them, roughly halving the desktop tool schema sent on every call, and about 2,100 tokens came off the fixed prompt. |
Opt-in caching (cache_ttl: auto) | Caches for an hour in interactive chats and five minutes for subagents and cron. Off unless you set it. |
| No surprise billing | Auto-routing never bills a provider you didn't pick, and provider_routing.models pins an OpenRouter provider per model. |
| Usage in the model picker | Shows how much of a subscription is left, when it resets and when a rate limit lifts, so you can switch before you hit the wall. |
| Bounded outage waits | Transient 503, 529 and timeout errors are retried in visible cycles (agent.auto_recovery_cycles, default 5) instead of ending the turn. |
Source: Hermes Agent v0.22.0 release notes, checked October 14, 2026.
What the lighter prompt saves
2,100 fewer tokens matter most on cache misses, and agents that wake on a timer miss often: in our measurements a lone wake-up found only 29% of its prompt in the cache. On an efficient model the saving is cents.
| 2,100 tokens on Claude Sonnet 5.5 | Saved |
|---|---|
| Per wake-up, cache miss ($2 per million) | $0.0042 |
| Per wake-up, cached ($0.10 per million) | $0.0002 |
| Per month, one wake-up every 30 minutes, 29% cached | $4.38 |
Our estimate from the fixed-prompt cut alone, at Anthropic's list prices; the halved tool schema on desktop saves more.
What hour-long caching costs
Anthropic charges twice the input rate to write a one-hour cache entry and 1.25 times for a five-minute one, so cache_ttl: auto pays off when you come back to a chat within the hour and costs more when you don't.
Custom providers and keys
A base_url you configure is now honored everywhere, and Anthropic and OpenAI keys are never sent to another host. Named custom providers also run on the Codex app-server, --provider chatgpt picks ChatGPT sign-in, and Nous accounts without credits can pick models billed to a subscription they already hold.
Hermes can borrow the logins of Codex CLI and Claude Code on the same machine. If you want those plans left alone, set auth.adopt_external_logins: false.
Upgrade notes that can bite
- Ended sessions inactive for 90 days are now deleted at startup. Pinned and open sessions are kept; set
sessions.auto_prune: falseto keep everything. - Chats no longer reset on an idle or daily timer. Use /new or /reset, and keep an eye on context growth in long chats, since every turn re-sends it.
- Source installs now follow tagged releases instead of every commit on main.
hermes update --set-channel mainswitches back. - The Windows desktop app is MSIX-only and needs Windows 11 22H2 or newer; the macOS app is Apple Silicon only. The installer scripts still work on Windows 10.
- Managed installs run on Python 3.14, and one messaging gateway per machine now serves every profile.
- Memory providers such as Honcho, Supermemory and Mem0 moved to catalog plugins. Existing users are migrated with their data.
If v0.22 costs you more
- Turned on
cache_ttl: autoand your chats sit idle for hours? Turn it off; the one-hour writes cost more than they save. - Trial model gone and a GPU with enough memory? Use a local model for routine work and keep a hosted model for the hard steps.
- Cron jobs and heartbeats on a frontier model? Move them to an efficient model. They miss the cache most and gain the least from a big model.
- Bills that swing month to month? A flat API caps them: Standard Compute is one option, below. Nous Portal makes more sense when you want its hosted tools, and a local model costs least if your hardware can run it.
The catch
None of v0.22's savings caps your bill. The free model is a one-time setup trial, local models are limited by your GPU, and a long agentic loop on a frontier model still costs whatever the provider charges per token.
Where Standard Compute fits
Hermes connects to OpenAI-compatible providers, so it runs on Standard Compute with standardcompute launch hermes, or base URL https://api.stdcmpt.com/v1 and model standardcompute. Plans start at $19 a month with a $20 compute budget, and requests stop at the budget instead of billing overage. Smart routing sends routine agent steps to efficient models, which suits heartbeats and cron jobs.
FAQ
What is the nous/welcome model in Hermes?
A free trial model that runs the setup chat the first time you open the desktop app on a computer, once per OS user. It needs no key or account. The CLI, TUI and messaging gateway don't get it.
How do I update to Hermes Agent v0.22?
Run hermes update on a CLI install, pull nousresearch/hermes-agent:stable for Docker or Hermes Cloud, and let the desktop app update itself. Moving to an older release is labelled a rollback and always asks first.
Does Hermes Agent v0.22 delete old sessions?
Yes, by default. Ended sessions inactive for 90 days are deleted at startup; pinned and open sessions are never touched. Set sessions.auto_prune: false before you update to keep everything.
Can Hermes v0.22 use my own OpenAI-compatible endpoint?
Yes. A configured base_url is honored everywhere in v0.22, and Anthropic and OpenAI keys are never sent to other hosts. Any OpenAI-compatible API works, including Standard Compute at https://api.stdcmpt.com/v1.