Best Models for Hermes Agent in 2026: Cost and Value
Hermes Agent still defaults to GLM-5.2. Verified September 2026 prices and coding scores show a model at a ninth of the cost now scoring higher.
Hermes Agent will drive more than 300 models, and that is exactly the problem. The software is free and open source. The model is the line item you pay for, and it is the one decision the setup wizard cannot make for you.
Hermes ships with an answer already selected. Its published model catalog, regenerated on September 3, 2026, sets z-ai/glm-5.2 as the default on both OpenRouter and Nous Portal. That default is now the wrong pick, and the gap is not small.
What is the best model for Hermes Agent?
For most people, GLM-5.3-Flash. It posts a composite coding score of 71.5 at a blended cost near $0.24 per million tokens, which works out to roughly six times the capability per dollar of anything else on the board. If you need the strongest reasoning available and can pay for it, use Claude Opus 5 at 78.0. If you want to spend nothing at all, Hermes ships nine free model ids in its catalog, and one of them is a free tier of the default. Everything below is how those shake out, and where a more specific pick beats all of them.
The rule that decides which models are eligible
Settle this before comparing anything, because it removes candidates that otherwise look fine on paper.
Nous states plainly that Hermes Agent requires at least 64,000 tokens of context for agent use with tools, and the wording in the documentation is a rejection, not a suggestion: smaller windows are refused at startup, because the system prompt, tool schemas, and running conversation state have to fit with room to work.
Two traps follow from that, and both are documented.
The first is that you cannot set context length through the OpenAI-compatible /v1/chat/completions API. It has to be configured on the server. On Ollama that means OLLAMA_CONTEXT_LENGTH=64000 ollama serve before the server starts, or a PARAMETER num_ctx 64000 line in a Modelfile. Nous calls this the single biggest source of confusion when connecting Ollama to tools like Hermes, and Ollama’s own default below 24GB of VRAM is 4,096 tokens, which is sixteen times too small.
The second is that parallel slots divide the window. Running llama.cpp with -c 64000 -np 4 gives each slot 16,000 tokens, below the per-session minimum, and the agent fails in a way that looks like a model problem rather than a config problem.
A model that benchmarks well at 8K is useless here. Check that number before you check anything else.
The value quadrant: capability against cost
Plot what a model can do against what it costs and the field sorts itself. Up and to the left is the sweet spot: strong capability, low price.
| Item | Blended cost, USD per million tokens (lower is better) | Composite coding score |
|---|---|---|
| Claude Opus 5 | $10 | 78 |
| Gemini 3.7 Flash | $1.5 | 76.1 |
| Claude Opus 4.8 | $10 | 74.3 |
| Grok 4.5 | $3 | 72.4 |
| Kimi K3 | $6 | 72 |
| GLM-5.3-Flash | $0.24 | 71.5 |
| Claude Sonnet 5 | $4 | 66.4 |
| GLM-5.2 | $2.15 | 46.5 |
Two things stand out. GLM-5.3-Flash is the point that breaks the pattern, holding a score within seven points of the best model on the chart at a fortieth of its price. And GLM-5.2, the model Hermes still selects for you by default, now sits at the bottom of the plot.
The default is stale, and here is the number
The July 2026 version of this guide recommended GLM-5.2 as the best-value pick, on a composite coding score of 68.8 that matched Gemini 3.1 Pro. That number no longer holds.
Re-reading the same source on September 4, 2026 returns 46.5 for GLM-5.2, with its intelligence composite falling from 51.1 to 34.8. This is worth stating carefully, because benchmark aggregates do drift and a single moved number is usually noise. It is not noise here. Every other model carried over from the July read came back unchanged: Claude Opus 4.8 at 74.3, Claude Fable 5 at 76.5, Gemini 3.1 Pro at 68.8, Qwen3.7 Max at 66.0, Kimi K2.7 Code at 60.8. GLM-5.2 moved alone, and it moved 22 points.
The practical consequence is that a fresh Hermes install still points at GLM-5.2, and Z.ai’s own successor is both cheaper and better. Changing it is one line.
Best value: GLM-5.3-Flash
Here is the pick to hand almost anyone starting out.
GLM-5.3-Flash is Z.ai’s MIT-licensed 320B-parameter mixture-of-experts model with 18B active parameters, released on August 26, 2026 with a 1,048,576-token context window. It runs $0.15 per million input tokens and $0.50 output, with cache reads at $0.03. On the blended measure used throughout this piece that is about $0.24, against $2.15 for GLM-5.2 and $10.00 for Opus 5.
It scores 71.5 on the composite coding index. Set those two numbers against each other and it delivers roughly 300 points of coding score per dollar of blended cost. The next best on that measure is Gemini 3.7 Flash at about 51, and Opus 5 at about 8. It is not a close contest, and it comfortably clears the 64K context rule with a window fifteen times larger than the floor.
Two honest caveats. First, “open weights” does not mean “runs on your laptop”: a 320B-total model is roughly 180GB at 4-bit, so self-hosting this one is a multi-GPU job. Second, Z.ai ran a launch promotion halving all three rates to $0.075, $0.25 and $0.015, scheduled to end on September 9, 2026. The figures above are the standard rates, so they hold after the promotion lapses.
Best raw performance: Claude Opus 5
When a task is genuinely hard, a long refactor across a repo or a multi-step research job with real branching, the cost of a wrong answer dwarfs the token bill. That is where you stop optimizing for price.
Claude Opus 5, released July 24, 2026, posts a 78.0 composite coding score, the highest of any model here, and it replaced Opus 4.8 at the identical $5 per million input and $25 output. That is roughly $10 blended: real money at volume, and trivial next to an hour spent debugging a plausible-looking wrong answer.
Claude Fable 5.1 shipped on September 1, 2026 at $10 and $50, with the cache-read rate cut to $0.25. Its predecessor Fable 5 scored 76.5, below Opus 5, and 5.1 does not yet carry a composite from the source used here. Until it does, Opus 5 is the defensible top pick rather than the most expensive one.
Best Gemini model for Hermes Agent
Gemini 3.7 Flash, and the reason is that the Flash tier stopped being the compromise option.
It scores 76.1 on the composite coding index, above Claude Opus 4.8 at 74.3, from a model priced at $0.75 input and $3.75 output, or $1.50 blended. That is a Flash-tier price attached to a near-frontier coding score, and it makes Gemini 3.7 Flash the strongest option in the field for anyone who wants one model to cover both the cheap chores and the hard jobs.
Google shipped Gemini 3.8 Flash on September 2, 2026 at the same $0.75 and $3.75. It has no composite score yet, so this guide cannot rank it. If it holds the 3.7 line it becomes the pick by default, since it costs the same.
Best Grok model for Hermes Agent
Grok 4.5, not the newer Grok 4.6, and this is the one recommendation here that runs against intuition.
Both models cost $2 per million input tokens and $6 output, an identical $3.00 blended. On the composite coding index read on the same day, Grok 4.5 scores 72.4 and Grok 4.6 scores 66.3. Same price, six points apart, in favour of the older release. Grok 4.6 does hold a higher general intelligence composite, 51.7 against 55.8, so the picture reverses on non-coding work.
For a Hermes install doing tool-calling and code, the version number is not the thing to sort on.
Best DeepSeek model for Hermes Agent
DeepSeek V4 Flash rather than V4 Pro, on both price and a billing detail that matters more for an agent than for a chatbot.
V4 Flash runs $0.44 input and $1.32 output at peak, about $0.66 blended, against V4 Pro at $1.32 and $3.96, or $1.98 blended. Both ship 1M-token context windows, so the 64K floor is a non-issue, and both support tool calls.
The detail: since August 16, 2026 DeepSeek bills by time of day, with off-peak rates at half the peak figures quoted above. Peak runs 01:00 to 04:00 and 06:00 to 10:00 UTC. An always-on agent with a scheduled heartbeat runs straight through both windows, so a Hermes deployment sees more of the peak rate than a human user would. Budget at peak and treat the off-peak half as upside.
Best small and cheapest options
The cheapest way to run Hermes is not a cheap model. It is a free one, and Hermes ships a list.
The published catalog carries nine ids flagged free, including z-ai/glm-5.2:free, minimax/minimax-m3:free, nvidia/nemotron-3.5-lightning:free, thinkingmachines/inkling:free and poolside/laguna-xs-2.1:free. That is a genuinely zero-cost path that does not require you to own a GPU, and the free tier of the catalog default is on it.
Below that, the cheapest paid model that clears the bar comfortably is GLM-5.3-Flash again, at about $0.24 blended. There is no meaningful saving in dropping to a weaker paid model from there.
| Item | Value |
|---|---|
| Claude Fable 5.1 | $20.00 |
| Claude Opus 5 | $10.00 |
| GPT-5.6 Sol | $8.00 |
| Kimi K3 | $6.00 |
| Claude Sonnet 5 | $4.00 |
| Grok 4.5 | $3.00 |
| GLM-5.2 | $2.15 |
| DeepSeek V4 Pro | $1.98 |
| Gemini 3.7 Flash | $1.50 |
| DeepSeek V4 Flash | $0.66 |
| GLM-5.3-Flash | $0.24 |
What Hermes Agent costs to run per month
Per-token rates are hard to feel. Here is the same choice expressed as a monthly bill.
Take a moderately active personal agent at 20 million input tokens and 4 million output tokens a month, with no credit for prompt caching. That budget is a stated assumption rather than a measurement, and a heavier or lighter install will scale roughly linearly from it.
| Model | Blended $/Mtok | Modelled monthly bill |
|---|---|---|
| GLM-5.3-Flash | $0.24 | $5.00 |
| DeepSeek V4 Flash | $0.66 | $14.08 |
| Gemini 3.7 Flash | $1.50 | $30.00 |
| Grok 4.5 | $3.00 | $64.00 |
| Claude Opus 5 | $10.00 | $200.00 |
| Claude Fable 5.1 | $20.00 | $400.00 |
The same agent, doing the same work, costs $5 a month or $400 a month depending on one config line. That is the spread the model decision actually controls, and it is why the default being 22 points out of date is worth five minutes of your attention.
This covers the token line only. For hosting, maintenance hours and the full first-year picture, see what a personal AI agent actually costs.
The model Nous tells you not to run
Worth knowing before you go looking: Nous Research recommends against running its own Hermes models inside Hermes Agent.
The Nous Portal documentation makes the Hermes 4 family available at heavily discounted rates and then says they are “not recommended for use inside Hermes Agent”, because “Hermes 4 is tuned for chat and reasoning, not the rapid-fire tool-calling loop the agent relies on”.
The same page names four picks by role: anthropic/claude-sonnet-4.6 as the best general-purpose agentic model, openai/gpt-5.5-pro for reasoning and tool calling, google/gemini-3-pro-preview for a large context window, and deepseek/deepseek-v4-pro as a cost-effective coder. Those are role assignments rather than a ranking, and the list has not tracked the last two months of releases, but it is the only first-party model guidance that exists.
Running Hermes on your own hardware
Some work should never leave your machine, and Hermes supports fully local inference through Ollama, vLLM, SGLang, llama.cpp and LM Studio with no API key.
The catch is that the 64,000-token floor is a hardware requirement, not a model one. Ollama defaults to 4,096 tokens on a card under 24GB, and a 64K cache costs several gigabytes on top of the weights, which is what decides whether a model actually fits. The full breakdown, with picks by card and the server flags tool calling needs, is in the guide to the best local models for Hermes Agent.
Which model for which job
The single-best-model question is the wrong one for an agent that does everything. Match the model to the task instead.
| The job | Run this | Why |
|---|---|---|
| Best all-round daily driver | GLM-5.3-Flash | 71.5 coding at about $0.24 blended, roughly 6x the next best value |
| Hardest coding and long refactors | Claude Opus 5 | 78.0 coding, the highest here, and strong long-horizon tool use |
| One model for cheap and hard work | Gemini 3.7 Flash | 76.1 coding at $1.50 blended, above Opus 4.8 at a seventh the cost |
| Best Grok option | Grok 4.5 | 72.4 against Grok 4.6 at 66.3, for the same $3.00 blended |
| Best DeepSeek option | DeepSeek V4 Flash | A third the price of V4 Pro, 1M context, tool calls |
| Spending nothing | A :free catalog id |
Nine free ids ship in the catalog, including the default |
| Fully local and private | Open weights via Ollama | Nothing leaves your box; needs a capable GPU and a 64K server config |
Routing instead of picking
The v0.21.0 “Pantheon” release, published August 31, 2026, added a model_overrides config block that patches a model’s context window, pricing or capabilities without waiting for a release. It also added six providers and a wave of catalog models including GLM-5.3-Flash itself.
That changes the shape of this decision. The older advice was to pick one model and re-pick later. With per-model overrides and a fuzzy /model picker, the practical setup is to keep a cheap model as the default and escalate specific jobs, rather than paying frontier rates on every chore.
The point of picking is that you can re-pick
The reason to sweat this decision less than it feels like you should: with Hermes, it is reversible. The whole design is that the model is a swappable input and the accumulated skills, memory and orchestration are what you keep. Start on GLM-5.3-Flash. If a task proves too hard, escalate that job to Opus 5 for a day. When the next cheap model lands, move the high-volume work over and lose nothing.
That is also where this meets the real cost of running an agent in production: the token bill is the volatile, negotiable line, and it moves every few weeks as new models land. The site’s model registry and leaderboard track those prices as they change, and the cost calculator turns a per-token rate into a per-task number for your own workload. Pick a model for today. Re-pick in a month. The harness does not care.
Frequently asked questions
Frequently asked questions
- What is the best model for Hermes Agent in 2026?
- GLM-5.3-Flash for most users: a 71.5 composite coding score at about $0.24 per million tokens blended, roughly six times the capability per dollar of the next model. Use Claude Opus 5, at 78.0, when a task is hard enough that a wrong answer costs more than the tokens. Both clear the 64,000-token context minimum Hermes requires for tool use.
- What is the default model in Hermes Agent, and should I change it?
- Hermes sets z-ai/glm-5.2 as the default on both OpenRouter and Nous Portal, per its published catalog dated September 3, 2026. Change it. GLM-5.2 was revised down from 68.8 to 46.5 on the composite coding index between July and September 2026, while its successor GLM-5.3-Flash scores 71.5 at roughly a ninth of the price.
- What is the best small model for Hermes Agent?
- Small is constrained by the 64,000-token context floor rather than by parameter count, and most small models fail that test before capability matters. The cheapest option that clears it is a free catalog id: Hermes ships nine, including a free tier of GLM-5.2 and Nemotron 3.5 Lightning. For paid, GLM-5.3-Flash at about $0.24 blended is cheaper than most small models.
- What is the best Grok model for Hermes Agent in 2026?
- Grok 4.5, not Grok 4.6. Both cost $2 per million input tokens and $6 output, an identical $3.00 blended, but Grok 4.5 scores 72.4 on the composite coding index against 66.3 for Grok 4.6. The newer model does lead on general intelligence, so the ranking reverses for non-coding work.
- What is the best Gemini model for Hermes Agent in 2026?
- Gemini 3.7 Flash, which scores 76.1 on the composite coding index, above Claude Opus 4.8, at $0.75 input and $3.75 output or about $1.50 blended. Gemini 3.8 Flash shipped on September 2, 2026 at the same price but carries no composite score yet, so it cannot be ranked here.
- How much does Hermes Agent cost to run per month?
- The software is free and MIT licensed, so the bill is tokens plus hosting. Modelled on 20 million input and 4 million output tokens a month, GLM-5.3-Flash lands near $5, DeepSeek V4 Flash near $14, Gemini 3.7 Flash near $30 and Claude Opus 5 near $200. The model choice moves the bill by a factor of forty.
- What is the cheapest Hermes Agent setup?
- A free catalog id, then local. Hermes ships nine free model ids including z-ai/glm-5.2:free and nvidia/nemotron-3.5-lightning:free, which costs nothing and needs no GPU. Running locally through Ollama costs only electricity but requires hardware that can hold a 64,000-token context, which rules out most laptops.
- What context window does Hermes Agent need?
- At least 64,000 tokens, per Nous documentation, and smaller windows are rejected at startup. Context length cannot be set through the OpenAI-compatible API, so it must be configured server-side: OLLAMA_CONTEXT_LENGTH=64000 on Ollama, or num_ctx in a Modelfile. Parallel slots divide the window, so llama.cpp with -c 64000 -np 4 leaves each session below the minimum.
- Can Hermes Agent run local models for free?
- Yes. Hermes supports local inference through Ollama, vLLM, SGLang, llama.cpp and LM Studio with no API key. Tool calling needs the right server flags: --enable-auto-tool-choice plus a tool-call parser on vLLM, and --jinja on llama.cpp. The catch is hardware, because a real 64K context on a frontier-class open model needs a capable GPU.
- Should I use Nous Research Hermes 4 models with Hermes Agent?
- No. Nous states in its own Nous Portal documentation that the Hermes 4 models are not recommended for use inside Hermes Agent, because Hermes 4 is tuned for chat and reasoning rather than the rapid-fire tool-calling loop the agent depends on. They are offered at a discount through the Portal, but the vendor advises against them for this job.
Sources
- Nous Research (2026). Hermes Agent documentation: LLM and model providers (context minimum, local backends, tool-calling flags). Re-verified 2026-09-04. https://nousresearch.github.io/hermes-agent/docs/integrations/providers
- Nous Research (2026). Hermes Agent documentation: Nous Portal (named model picks, Hermes 4 anti-recommendation). Verified 2026-09-04. https://nousresearch.github.io/hermes-agent/docs/integrations/nous-portal
- Nous Research (2026). Hermes Agent model catalog (JSON manifest, updated 2026-09-03; catalog default, recommended flag, free ids). https://nousresearch.github.io/hermes-agent/docs/api/model-catalog.json
- NousResearch (2026). Hermes Agent releases (v0.21.0 “Pantheon”, tag v2026.8.31, published 2026-08-31;
model_overrides). https://github.com/NousResearch/hermes-agent/releases - Anthropic (2026). Claude pricing and the Claude Opus 5 announcement. https://claude.com/pricing
- Z.ai (2026). GLM-5.3-Flash model page (rates, context window, MIT weights). https://openrouter.ai/z-ai/glm-5.3-flash
- DeepSeek (2026). API pricing (official documentation, time-of-day billing). https://api-docs.deepseek.com/quick_start/pricing
- Artificial Analysis (2026). Model comparison: coding and intelligence indices, aggregated via Price Per Token. All rows re-read in one pass 2026-09-04. https://artificialanalysis.ai/models
- Capital & Compute. AI model registry (per-token rates, verified September 2026). https://capitalandcompute.net/ai-models/
Prices and capability scores were re-verified against the sources above on September 4, 2026, replacing a July 5, 2026 reading in which GLM-5.2 scored 68.8 rather than 46.5. Per-token rates and benchmark composites change frequently. See the editorial standards.