# Hermes Agent Cost 2026: Token Usage, Models and Savings

We ran the same four tasks through Hermes Agent 18 times on six setups and logged every token. Running it through SayGM cost about half of buying the same model direct, a session averaged $0.004 to $0.10, and one setting cut the bill by about 60%. Here's where the tokens go and how to spend less.

_Source: https://saygm.com/blog/hermes-agent-cost_
**Measured 10 November 2026.** Hermes Agent is free; running it costs about $8 to $37 a month at 10 sessions a day on a $7 server, from Qwen3.8 27B at the low end to Claude Opus 5.5 at the top. Buying the model through SayGM roughly halves the token bill, and limiting Hermes to the tools you need cuts it by another 60%.

**Monthly cost, all in (10 sessions a day, on a $7 server), checked 10 November 2026**

| Model | Total a month | Performance |
| --- | --- | --- |
| [Qwen3.8 27B](https://saygm.com/models/qwen3.8-27b) | $8 | ★★★★☆ |
| [Claude Sonnet 5](https://saygm.com/models/claude-sonnet-5), tools limited | $15 | ★★★★★ |
| Claude Haiku 4.5 | $16 | ★★☆☆☆ |
| [GLM 5.3](https://saygm.com/models/glm-5.3) | $28 | ★★★☆☆ |
| Claude Sonnet 5 | $28 | ★★★★☆ |
| [Claude Opus 5.5](https://saygm.com/models/claude-opus-5-5) | $37 | ★★★★★ |

*Stars: 5, minus 1 for each wrong web answer, minus 1 if reasoning showed up in replies, minus 1 for a data error.*

*Method: Hermes Agent v0.20.0 (the current release is v0.23.0) on a 4 GB Hetzner VPS, three fresh sessions per setup, each running the same four tasks (a short explanation, a file edit, a web lookup and a data job that writes and runs a Python script). Every setup finished all four tasks. Cost is the mean per session, from Hermes' per-call token log times SayGM's rates on 10 November 2026, times 300 sessions a month, plus $7 for the server.*

**Your bill will be different.** These are four short, fixed tasks run the same way on every setup, so they show how the models and settings compare. They won't predict your own bill. Real use can cost far more or far less: a long coding session, a big repository, large web pages or files in the context, or an agent left running on a schedule can use tens or hundreds of times the tokens of one of our sessions. Use the per-session figures to compare options, then check your own spend after a week.

## Table of Contents

- [Monthly cost breakdown: tokens, server and direct prices](#monthly-cost-breakdown-tokens-server-and-direct-prices)
- [How much does Hermes Agent cost to run?](#how-much-does-hermes-agent-cost-to-run)
- [Hermes Agent token usage: where the tokens go](#hermes-agent-token-usage-where-the-tokens-go)
- [How to reduce Hermes Agent token usage](#how-to-reduce-hermes-agent-token-usage)
- [What is the cheapest model for Hermes Agent?](#what-is-the-cheapest-model-for-hermes-agent)
- [Is Hermes cheaper than a ChatGPT or Claude subscription?](#is-hermes-cheaper-than-a-chatgpt-or-claude-subscription)
- [Running Hermes on sensitive work](#running-hermes-on-sensitive-work)
- [Hermes Agent cost FAQ](#hermes-agent-cost-faq)
- [TL;DR](#tldr)
- [About SayGM](#about-saygm)

## Monthly cost breakdown: tokens, server and direct prices

At 300 of our test sessions a month on a $7 server, as of 10 November 2026. Your own usage will move these numbers a long way in either direction; the gaps between rows are the useful part.

| Model | Tokens, SayGM | Server (VPS) | Total a month, SayGM | Total a month, direct | Performance |
| --- | --- | --- | --- | --- | --- |
| Qwen3.8 27B | $1.20 | $7 | $8 | $18 | ★★★★☆ |
| Claude Sonnet 5, tools limited | $8.40 | $7 | $15 | $23 | ★★★★★ |
| Claude Haiku 4.5 | $8.70 | $7 | $16 | $24 | ★★☆☆☆ |
| GLM 5.3 | $20.70 | $7 | $28 | $47 | ★★★☆☆ |
| Claude Sonnet 5 | $21.30 | $7 | $28 | $48 | ★★★★☆ |
| Claude Opus 5.5 | $30.00 | $7 | $37 | $65 | ★★★★★ |

On a small VPS, the server is most of the bill for the cheaper models. On Hippius Confidential Compute at about $36 a month, add about $29 to each total.

*Performance uses the same star rule as the table at the top. Direct means Anthropic's list price for Claude, Z.ai's for GLM 5.3 and Alibaba's for Qwen3.8 27B.*

## How much does Hermes Agent cost to run?

As of November 2026, Hermes Agent costs nothing to install, and running it typically costs a few dollars to about $30 a month in model tokens plus about $7 to $36 a month for a server, depending on the model you choose. [Hermes Agent](https://hermes-agent.nousresearch.com/) is Nous Research's open-source agent, released under the MIT licence, so there's no licence fee. The bill has three parts.

**Hosting.** Hermes needs somewhere to run all day if you use it from Telegram, Discord or a cron job. Our test box was a Hetzner VPS with 4 GB of RAM. Hetzner's cheapest shared plan, the CX23 with 2 vCPUs and 4 GB, lists at €5.99 a month excluding VAT with IPv4 included, though it showed as unavailable when we checked on 9 November. Want your agent's keys and prompts kept private? Run Hermes Agent on [Hippius Confidential Compute](https://hippius.com/confidential-compute) for about $36 a month. Its memory is encrypted, so even the server's operator can't read it.

![Hippius Confidential Compute page, VMs with memory encrypted by the processor](https://assets.saygm.com/blog/2026/11/4599e567-a67f-4d0f-9a22-71d513d50511-hippius-confidential-compute-page-november-2026-1456x829.jpg)

**Model tokens.** This is the part that varies. Every Hermes call resends its system prompt, tool definitions and conversation history, so the price per token and the size of that prompt decide the bill. Our measured sessions averaged $0.004 to $0.10, with single runs up to $0.15.

**Optional services.** Web search, browser use and image generation can need their own keys, or you can pay per use through the Tool Gateway on a [Nous Portal](https://portal.nousresearch.com/) plan. Our test used Hermes' built-in search, which cost nothing extra.

**Local models.** Running a model on your own machine costs nothing per token, but Hermes needs at least a 64K context window, which rules out the smallest models on low-memory hardware. The docs suggest a 16 GB machine for a comfortable 128K context.

## Hermes Agent token usage: where the tokens go

As of Hermes Agent v0.20.0, tested in November 2026, every call carries a fixed prompt of about 30,600 tokens on Claude with the default tools, before your message or any history is added. That's the system prompt plus the definitions of 31 tools, sent again on every call.

![Hermes Agent fixed prompt per call by setup: about 30,600 tokens on Claude Sonnet 5 with default tools, 6,600 with web, terminal and file tools only, November 2026](https://assets.saygm.com/blog/2026/11/29692f57-096b-4dfc-918a-e21f959c6513-hermes-agent-tokens-per-call-november-2026-3560x1476.png)

| Setup | Tokens in the first call |
| --- | --- |
| Claude Sonnet 5, default tools | about 30,620 |
| Claude Opus 5.5, default tools | about 30,560 |
| Claude Haiku 4.5, default tools | about 22,800 |
| Qwen3.8 27B, default tools | about 21,040 |
| GLM 5.3, default tools | about 20,010 |
| Claude Sonnet 5, web, terminal and file tools only | about 6,600 |

A four-task session made 9 to 15 calls, so that fixed prompt was sent 9 to 15 times. Two things keep it from dominating the bill.

**Prompt caching works through SayGM.** Hermes adds Anthropic cache markers on Claude requests to Anthropic-compatible endpoints, and its startup banner read "Prompt caching: ENABLED (Anthropic-compatible endpoint, 5m TTL)". In a five-turn test conversation on Sonnet 5, 99 to 100% of input after the first call came from cache. On SayGM, cached input costs about 90% less than fresh input: $0.103 per million tokens against $1.03 on Sonnet 5.

The cache lasts five minutes and is shared between sessions with the same prompt, so a session started within five minutes of the last one skips the first-call write.

**A cold start costs extra.** When nothing is cached, the first call pays to write the whole prefix. On Opus 5.5 that made the first task cost $0.08 in a cold session against $0.004 in a warm one.

Long sessions add history on top. In a nine-task session that pasted large log extracts into the conversation, the prompt reached about 130,000 tokens and the session cost about $0.27 on Sonnet 5.

One caution on older numbers: a widely shared GitHub analysis put Hermes' fixed overhead at about 13,900 tokens per call, but that was on v0.6.0 and the issue is closed. On v0.20.0 we measured more than twice that with the default tools.

## How to reduce Hermes Agent token usage

Easiest savings first.

**1. Use SayGM: about 48% off every token.** The same sessions, priced at each model maker's own rates and at SayGM's, on 10 November 2026:

| Model | Per session, direct | Per session, SayGM | Per month, direct to SayGM |
| --- | --- | --- | --- |
| Claude Opus 5.5 | $0.194 | $0.100 | $58 to $30 |
| Claude Sonnet 5 | $0.137 | $0.071 | $41 to $21 |
| GLM 5.3 | $0.134 | $0.069 | $40 to $21 |
| Claude Haiku 4.5 | $0.057 | $0.029 | $17 to $9 |
| Qwen3.8 27B | $0.035 | $0.004 | $10.50 to $1.20 |

Claude is 48.5% off Anthropic's list price on SayGM, GLM 5.3 about 49% off Z.ai's, and Qwen3.8 27B about 89% off Alibaba's. To set it up, [use Hermes Agent with SayGM](https://saygm.com/integrations/hermes).

**2. Limit the tools to what the job needs (about 60% saved).** Restricting Sonnet 5 to web, terminal and file tools cut the fixed prompt from about 30,600 to 6,600 tokens and the session cost from $0.071 to $0.028, and all three runs still got the web answer right. Combined with SayGM, a Sonnet 5 session went from $0.137 bought direct on the default setup to $0.028, about 80% less. Pick toolsets per session:

```bash
hermes chat --toolsets "web,terminal,file"
```

Or turn off the ones you never use, everywhere:

```yaml
# ~/.hermes/config.yaml
agent:
  disabled_toolsets:
    - browser
    - memory
```

**3. Choose the model for the job (up to 95% saved).** Qwen3.8 27B finished every task for $0.004 a session against $0.071 on Sonnet 5. See [the next section](#what-is-the-cheapest-model-for-hermes-agent) for when that isn't enough.

**4. Keep the cache warm (about 90% off cached input).** Caching is on by default for Claude through SayGM. Back-to-back sessions reuse it; a gap of more than five minutes means paying to write it again. If you type into Hermes with longer breaks, `cache_ttl: "auto"` uses a one-hour cache for human-paced sessions, which costs more per write and pays off only when your turns are more than five minutes apart:

```yaml
prompt_caching:
  cache_ttl: "auto"
```

**5. Move compression to a cheap model (small saving).** When a session gets long, Hermes summarises the middle of it. With the summary on Qwen3.8 27B, each summary cost $0.001 to $0.003 against $0.013 to $0.016 on Sonnet 5, so about $0.01 saved per compression, on sessions costing about $0.30. Qwen took longer to write it (36 to 78 seconds against 18 to 24). Pick a summariser with a context window at least as large as your compression threshold; a smaller one has caused compression loops in the past.

```yaml
auxiliary:
  compression:
    base_url: https://api.saygm.com/v1
    api_key: ${SAYGM_API_KEY}
    model: qwen3.8-27b
```

Compression itself didn't make our long sessions cheaper. Rewriting the history invalidates the cache, so the next call pays to write it again, and the sessions that compressed at about 86,000 tokens cost the same as or slightly more than the one that never did. Compression keeps a very long prompt from growing; at these sizes it isn't a saving.

**6. Prune bulky tool output (not measured).** Hermes' docs describe `proactive_prune_tokens`, off by default, which trims old oversized tool results without a model call once resent history passes a set size. The docs suggest 48,000 as a starting point. We didn't test it.

## What is the cheapest model for Hermes Agent?

As of November 2026, the cheapest model that finished every task in our Hermes test was Qwen3.8 27B through SayGM, at about $0.004 per four-task session. It also gave the right answer on the web lookup in all three runs.

It isn't flawless. Qwen and GLM 5.3 both printed their reasoning into the final answer ("The user is asking a simple question..."), which adds output tokens and clutter that the Claude models didn't have. On the long-session test, Qwen was also slower as a summariser.

| Model | Cost per session | Web answer right | Notes |
| --- | --- | --- | --- |
| Qwen3.8 27B | $0.004 | 3 of 3 | Reasoning shows in replies |
| Claude Haiku 4.5 | $0.029 | 1 of 3 | Vague or out-of-date web answers; one data table missing its total row |
| GLM 5.3 | $0.069 | 2 of 3 | Reasoning shows in replies |
| Claude Sonnet 5 | $0.071 | 2 of 3 | 3 of 3 with tools limited |
| Claude Opus 5.5 | $0.100 | 3 of 3 | Fewest calls per session |

The web lookup is where the models split. Hermes' default search returns search results only, so models that didn't open the source page sometimes answered from snippets, and that's where the wrong answers came from. Opus 5.5 cost about 1.4 times Sonnet 5 per session even though its tokens cost twice as much, because it needed fewer calls (10.3 against 12.3 on average). Three runs per model is a small sample, so treat that as a lead. Claude Haiku 5.5, the newer Haiku, costs a tenth of Haiku 4.5 per token on SayGM ($0.0515 per million input tokens against $0.515); we didn't test it in Hermes.

A cheap model isn't enough when the job is long, open-ended or hard to check: a refactor across many files, research where a wrong answer is costly, or a job you'll run unattended. That's where Sonnet or Opus earns its price.

## Is Hermes cheaper than a ChatGPT or Claude subscription?

For most light and moderate use, yes. At our measured rates, 300 short sessions a month cost $1.20 on Qwen3.8 27B, $8.40 on Sonnet 5 with tools limited and $30 on Opus 5.5, against $20 a month for a ChatGPT Plus or Claude Pro plan. Add about $7 to $36 for hosting.

The comparison isn't one-sided:

- **Free and local models.** Nous Portal's free tier gives access to free models only, with no hosted tools. A local model costs nothing per token if your hardware can hold a 64K context. Both are the cheapest routes when the model is good enough.
- **A ChatGPT plan.** Hermes can run on a ChatGPT account through its OpenAI Codex provider, and Nous Portal can connect an eligible ChatGPT account for inference without a paid Nous plan. If you already pay for ChatGPT, that may cost you nothing extra until you hit its limits.
- **Nous Portal plans.** Plus is $20 a month for $22 of credits, Super $100 for $110 and Ultra $200 for $220, with the Tool Gateway's web, browser and image tools charged per use from the same balance. Worth it if you want hosted tools and one bill.
- **Per-token through a gateway.** On SayGM, Claude through Hermes costs 51.5% of Anthropic's list price. The same Sonnet 5 sessions would have cost $0.137 each at Anthropic's list rates, against $0.071 on SayGM. OpenRouter passes the provider's rate through and adds a 5.5% fee on credit purchases; we set the two out in [SayGM vs OpenRouter](https://saygm.com/vs/openrouter).

Running Claude Code instead? See [Claude Code pricing](https://saygm.com/blog/claude-code-pricing) for its plans and per-token costs.

## Running Hermes on sensitive work

An agent sees what you give it: files, messages, keys. Two layers decide who else can see them.

On SayGM, every request passes through a gateway sealed in an Intel TDX enclave, which its operators can't read. On a Claude, GPT or other frontier model, the model provider still receives the prompt and processes it under its own API terms. On a sealed open-weight route, the prompt stays inside hardware-verified execution end to end, so nobody in the chain can read it.

Sealed routes cost more than standard ones. [Qwen3.8 27B (TEE)](https://saygm.com/models/qwen3.8-27b-tee) is $0.19 per million input tokens and $1.46 output on 10 November 2026, about four to five times the standard Qwen route, and GLM 5.3 (TEE) is $1.23 and $3.85 against $0.72 and $2.26. We didn't run the task set on a sealed route, so we can't give a measured session cost for one yet. To set either up, [use Hermes Agent with SayGM](https://saygm.com/integrations/hermes).

## Hermes Agent cost FAQ

### How much does Hermes Agent cost?

As of November 2026, Hermes Agent is free to install, and in our test a four-task session cost $0.004 on Qwen3.8 27B, $0.07 on Claude Sonnet 5 and $0.10 on Claude Opus 5.5 through SayGM. At 10 sessions a day that's about $1.20, $21 or $30 a month in tokens, plus about $7 to $36 a month for a server.

### Is Hermes Agent free?

Yes. As of November 2026, Hermes Agent is free and open source under the MIT licence. You pay for the model it calls, unless you use free or local models, and for hosting if you run it on a server.

### Why does Hermes use so many tokens?

Because every call resends a fixed prompt: about 30,600 tokens with the default tools on Claude in Hermes v0.20.0, tested November 2026, made up of the system prompt and 31 tool definitions. A session makes 9 to 15 calls, and the conversation history rides on top. Prompt caching makes most of that resent prompt cheap.

### How do I reduce token usage?

As of November 2026, the easiest saving is buying the model through SayGM, which roughly halved the token bill in our test, from $0.137 to $0.071 per Sonnet 5 session. Next, limit Hermes to the tools a job needs: --toolsets \"web,terminal,file\" cut the fixed prompt to about 6,600 tokens and a Sonnet 5 session's cost by about 60% more. After that, choose a cheaper model and keep sessions close together so the cache stays warm.

### What is the cheapest model for Hermes?

In our November 2026 test, Qwen3.8 27B was the cheapest model that finished every task, at about $0.004 per four-task session through SayGM. It prints its reasoning into replies, so for long or hard-to-check work, Claude Sonnet 5 with limited tools at $0.028 a session is the safer low-cost choice.

### Can I run Hermes on free or local models?

Yes. Nous Portal's free tier covers free models, and Hermes runs local models through any OpenAI-compatible server. The model needs at least a 64K context window, so on an 8 GB machine Hermes' docs suggest a smaller model with a compressed KV cache.

### Do I need a Nous Portal subscription?

No. As of November 2026, Hermes works with any supported provider and your own keys for tools. A Nous Portal plan, from $20 a month for $22 of credits, adds 200+ models and hosted tools billed per use from one balance.

### Does prompt caching work with custom providers?

For Claude on an Anthropic-compatible endpoint, yes: Hermes v0.20.0 adds cache markers there, and in our November 2026 test through SayGM, 99 to 100% of input after the first call was served from cache. Hermes' docs list only the native Anthropic API, OpenRouter and Nous Portal, so check your startup banner for "Prompt caching: ENABLED" on other providers.

## TL;DR

Hermes Agent is free; you pay for hosting and model tokens. In our measured test through SayGM, a four-task session cost $0.004 on Qwen3.8 27B, $0.07 on Claude Sonnet 5 and $0.10 on Claude Opus 5.5. Buying the model through SayGM roughly halved the token bill, and limiting the tools cut a Sonnet session by about 60% more. Prompt caching worked through SayGM without any setup. For the wider price picture, see the [cheapest LLM API](https://saygm.com/blog/cheapest-llm-api) comparison and how SOMA handles [AI agent costs](https://saygm.com/blog/how-soma-cuts-ai-agent-costs-with-saygm).

## About SayGM

SayGM is a drop-in inference gateway for teams who don't want to just take a company's word that their prompts are private. Route to a confidential model and your prompt is sealed inside hardware-verified execution, where not even SayGM can see it. That's not a policy, it's provable. Swap in your existing OpenAI, Anthropic, or Gemini code and you're covered in minutes, at transparent, published rates with no hidden markup.

Say gm to AI at [saygm.com](https://saygm.com).

[Website](https://saygm.com) | [Twitter](https://x.com/say_gm_) | [Discord](https://discord.com/channels/799672011265015819/1343950080465698836) | [Blog](https://saygm.com/blog) | [Medium](https://medium.com/@say_gm_) | [Docs](https://docs.saygm.com/)