Cost & models
Alternatives, split by who measured what
Four terminal agents are compared here, and the table is split by evidence: the Hermes Agent column is measured on this machine (install footprint, per-task cost on DeepSeek, the failure modes we hit), while Claude Code, OpenCode and Codex CLI are described only from their official documentation, fetched on 2026-11-22, and marked as not run. Their install commands and billing models are facts; their real cost per task on your repo is not something this page can tell you, because we never installed them.
People search “hermes agent alternatives” when they are deciding whether to switch. The honest answer has two halves with different reliability, so the table below is split by who measured what: one column we ran, three columns we only read.
The rule this page follows
A number we did not produce is labelled as documentation, not as a result. That is the whole editorial policy here, and it has a cost: the competitor columns are thinner than a marketing page would be. It also has a benefit: nothing on this page will be wrong in the way “benchmarks” usually are.
The three other columns were read on 2026-11-22 from docs.claude.com, opencode.ai/docs and github.com/openai/codex. None of them was installed, authenticated or run. If a sentence below about those tools is wrong, it is because their docs changed, not because we tested it.
How you install and how you pay
These are the two questions that decide most evaluations, and they are also the two the vendors document precisely.
| Agent · evidence | Install | How you pay |
|---|---|---|
| Hermes Agent · measured | uv tool on Windows; three executables, 116 MB |
Per token, on your provider’s API (DeepSeek here) |
| Claude Code · docs | native installer, Homebrew, WinGet, apt | Claude subscription or Console account; third-party providers allowed |
| OpenCode · docs | npm, brew, bun, mise — no Windows package managers | Bring your own provider, or a $10/month plan |
| Codex CLI · docs | native installer, npm, brew cask | ChatGPT plan, or an API key |
The exact commands, as each vendor writes them on 2026-11-22 (only the first was run here):
Hermes uv tool install <package>
Claude Code curl -fsSL https://claude.ai/install.sh | bash
OpenCode npm install -g @opencode/cli
Codex CLI curl -fsSL https://chatgpt.com/codex/install.sh | sh
Read the first column as the evidence tag: measured is our run, docs is the vendor page fetched 2026-11-22 and never executed here.
Two structural differences fall out of that table, and neither is a quality judgement:
- Subscription versus per-token. Three of the four can be paid for as a plan; Hermes cannot, because it is a client for a model API. A plan caps the bill; a price list makes the bill computable. Which is better depends on whether your usage is bursty.
- Which one you can install on Windows at all. OpenCode’s own docs say Windows package managers are not supported — that is a fact about deployment, not about the product.
What a task actually costs, in the one column we can compute
For Hermes on DeepSeek, the marginal cost is a multiplication: tokens in, cache-hit rate, tokens out, times the rate on the provider page. Using one fixed workload — 200,000 input tokens, 90% cache hit, 20,000 output tokens — the numbers on this machine are:
deepseek-flash, off-peak $0.0155
deepseek-v4-pro, peak $0.1135 7.3x
Those two come from prices.ts, which is read from the provider’s own rate page and dated. The off-peak multiplier is 2×, not 7.3× — the 7.3× is cheapest configuration versus priciest on the same task, which is a different claim and the one people usually mistype.
For the other three, this page can tell you the billing model and not the number, because the number needs your token counts and your plan. Anyone publishing a per-task dollar figure for a subscription product without those is inventing the denominator.
The single most-searched pairing — Hermes Agent against Claude Code — is split the same way on its own page, with their pricing page quoted as unreadable on the day we tried.
What you give up if you switch away
Not “features” — the four things we can point at with a file or a command:
| Thing | Where it is | After a switch |
|---|---|---|
| Recomputable per-token cost | prices.ts + your provider bill |
Subscription products do not expose unit prices; the multiplication disappears |
| Three separate stores | %LOCALAPPDATA%\hermes, ~/.hermes, per-profile dirs |
Each product has its own layout; migration untested |
| Error index by verbatim text | /symptoms/ — 23 entries, 8 literal |
Rebuilding it is the expensive part, not installing |
| Gates that decide “done” | check.mjs, postbuild.mjs, deploy.sh |
Written against this stack’s artifacts; they do not travel |
That list is why “just try the other one for a week” is cheaper than it sounds for a hobby project and much more expensive for a working setup.
How you know it worked
If you are evaluating rather than reading, the only number worth trusting is the one you produce. Run the same task in two agents, then take tokens from your own provider billing page — not from a calculator, not from this page — and multiply by the rate you actually pay. Three fields decide it: input tokens, cache-hit tokens, output tokens.
The check that catches self-deception: run the same task twice in the same agent on different days. If your cost per task moves more than the price list says it should, you are measuring your prompt, not the vendor.
How to undo it
Program and state are separate on this stack: the install directory is disposable, the state is not. On Windows it is %LOCALAPPDATA%\hermes; under WSL it is ~/.hermes; named profiles keep their own memory and session databases underneath. Removing the executable leaves all of it in place.
We have run the backup and the upgrade path, and we have not run an uninstall. So treat “uninstall is clean” as unverified — the safe order is: copy the state directory somewhere first, remove the program, then delete the copy once the replacement works.
What we did not run
The whole competitor half. No install, no login, no task, no cost measurement for Claude Code, OpenCode or Codex CLI. Their columns are what their documentation says on 2026-11-22, and nothing more.
We also did not test cross-tool migration in either direction — not their state into ours, not ours into theirs. If you need that answered before switching, this page cannot answer it, and neither can a table.
Still stuck on an error?
The line your terminal printed is often the fastest way in. 25 recorded errors and symptoms in 7 areas — 8 of them quoted verbatim, each pointing at the page that fixes it.
Shell & paths · MySQL · TLS & certificates · CSS & layout · API & keys · Agent memory · Cost & choice
Related
Questions people actually ask
Why not run all four and compare them head to head?
Because we do not have that setup. Our working stack is Hermes Agent on DeepSeek; installing three more agents and driving them through the same task would be a different project, and a comparison written without it would be a table of other people's claims with our name on top.
Then what is the doc-sourced column worth?
It answers the question people actually search: how do you pay, how do you install it, does it need a subscription. Those are verifiable facts on the vendor page, and we date them. What it cannot answer is "which is cheaper for my repo" — for that you need your own token counts.
Which one is cheapest?
Unknown, and anyone who tells you otherwise without your token counts is guessing. The only defensible comparison is marginal cost per task: tokens in, cache-hit rate, tokens out, times the rate you actually pay. We compute that only for the stack we ran.
Does Hermes Agent need a subscription?
No. It calls a model API directly, so the bill is per token on your provider account. That is also the reason the cost is computable here and not for the subscription products: we can read every unit price.
What happens to my state if I switch?
Program and state are separate: the install directory is disposable, the state under %LOCALAPPDATA%\\hermes (Windows) or ~/.hermes (WSL) is not — memory, skills, session databases. Deleting the program does not delete the state, and we have not tested a migration into any of the other three.
Sources
- Claude Code documentation (overview)
- OpenCode documentation
- openai/codex on GitHub
- Hermes Agent documentation
Last checked: 2026-11-22 · Hermes column measured on this machine (installer run 2026-11-19; costs read from src/data/prices.ts captured 2026-11-19 and from the local usage ledger: 1,076 calls / 99.1% cache hit / $1.11 off-peak over 30 days). Claude Code, OpenCode and Codex CLI columns come from their official docs, fetched 2026-11-22, and were not installed or run — every cell in those columns is documentation, not measurement.