Hermes Agent vs Claude Code, split by evidence
Searches for "hermes agent vs claude code" want a verdict. What they can get here is one measured half, one documented half, and an explicit line between them — because a comparison that mixes the two is how the genre us
Searches for "hermes agent vs claude code" want a verdict. What they can get here is one measured half, one documented half, and an explicit line between them — because a comparison that mixes the two is how the genre usually goes wrong.
The rule: two evidence levels, never one column
Measured means we ran it on this machine and can rerun it. Documented means the vendor's own page says so on a stated date. They are not interchangeable, and the one thing this post will not do is average them into a single score.
The Claude Code column was read on 2026-11-22 from docs.claude.com. It was not installed, not logged in, and no task was run on it.
If you are still at the wider question — which agents exist at all, and how each is paid for — that comparison is one page over, with the same split.
Install, and what lands on disk
| Hermes Agent · measured | Claude Code · documented | |
|---|---|---|
| Install | uv tool install |
native installer (claude.ai/install.sh), WinGet, Homebrew casks, apt/dnf/apk |
| What you get | three executables and a 116 MB directory | a CLI plus IDE extensions, a desktop app and a web surface |
| Updates | you re-run the upgrade path; we back up before it | native installs update in the background; Homebrew and WinGet installs do not |
| Windows note | installs natively; we also run the WSL backend | Git for Windows recommended on native Windows so the Bash tool has a shell; WSL needs it not |
The Claude Code row is more than we can verify: "native installs update in the background" is a behaviour we have not watched, and it is also the kind of behaviour that changes what "which version am I on" means.
How each one charges you
This is the structural difference, and it is not a quality difference:
- Hermes Agent is a client for a model API. You pay your provider per token, and every unit price is on a page you can read. That makes cost a multiplication you can do before you spend.
- Claude Code is billed through a Claude subscription or an Anthropic Console account. A plan caps the bill; it does not expose a per-token number to multiply.
Which is better depends on whether your usage is bursty (a cap helps) or steady and measurable (a rate helps). And on 2026-11-22 the second half of that sentence could not be checked: docs.claude.com/en/docs/about-claude/pricing answered "App unavailable in region". So this post contains no Claude Code price at all — not because it is unflattering, but because we could not read it.
One task, two configurations, our numbers
Fixed workload: 200,000 input tokens, 90% cache hit, 20,000 output tokens.
deepseek-flash, off-peak $0.0155
deepseek-v4-pro, peak $0.1135 7.3x
The 7.3× is cheapest configuration versus priciest on the same task. The peak-versus-off-peak multiplier is 2× — a different number that gets quoted as 7.3× often enough to be worth repeating.
Those rates come from a price file transcribed from the provider's own page with a capture date. The cache-hit rate is ours, not a vendor claim: 99.1% over 1,076 calls in a 30-day window, which is what makes the cheap configuration cheap in practice rather than in theory.
Three questions that decide it
Not a score — the three forks where the two products actually diverge:
- Do you need to know the cost before you spend, or only after? Per-token pricing answers "what will this task cost" before the run. A subscription answers "what did this month cost" after it.
- Is your work Windows-native? Both document Windows. Hermes' install footprint is one directory you can delete; Claude Code's docs recommend an extra dependency (Git for Windows) for its shell tool on native Windows.
- Do you want the model to be your choice? Hermes is provider-agnostic by construction — the bill is the provider's. Claude Code's docs say its terminal and IDE surfaces also accept third-party providers, which narrows this difference but does not remove it.
What this post will not tell you
Which one writes better code, handles a large repo better, or recovers better from a failed edit. All three need the same task run in both, and we have not installed Claude Code. If a page gives you those without running both, it is reporting someone else's numbers with its own name on top.
We also could not read their pricing page from this region on 2026-11-22, so any cost comparison that includes a subscription figure is out of scope here by construction, not by choice.
How you know it worked
For the reader this is a protocol, not a claim: run the same task in both, then take tokens from your own provider billing page and multiply by the rate you pay. The check that catches self-deception is to run the same task twice in the same agent on different days — if your per-task cost moves more than the rate list allows, you are measuring your prompt, not the vendor.
For the subscription side, the equivalent check is the cap: log used-vs-allowed before and after a fixed task, and see whether the task moved it by what you expected. We have not run that check on Claude Code.
How to undo it
Pieces, in the order that does not lose data:
- Keep the state, remove the program. On this stack the install directory is disposable;
%LOCALAPPDATA%\hermes(Windows) or~/.hermes(WSL) is not — memory, skills and per-profile session databases live there. - Copy the state directory before touching anything, and only delete the copy once the replacement is working.
- Revert the model configuration separately from the install: a client pointed at a dead endpoint looks exactly like a broken agent.
We have run backup and upgrade on this stack. We have not run an uninstall, and we have not tested migration into Claude Code. Both are statements about our evidence, not about the products.
What we did not test
Everything on the Claude Code side of the table: install, login, tool use, cost, and the background-update behaviour its docs describe. Also untested: cross-tool migration in either direction, and any claim about output quality on either side.
Originally published at agenthardway.com/cost/hermes-agent-vs-claude-code. The split-by-evidence rule this post follows is written up at agenthardway.com/method.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.