Best Models for Hermes Agent: Speed, Cost, Value
Hermes Agent runs 300-plus models, so which one should you actually run? A grounded 2026 guide to the best pick for performance, cost, and value.
By Capital & Compute
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. Point it at the wrong model and you either burn money on a frontier API for chores a cheap model would nail, or you starve a hard task with something that cannot hold a tool chain together.
So this is the picker. Verified per-token prices, independent capability scores, and one hard constraint most guides skip: Hermes needs a model with at least a 64,000-token context window before its tools work reliably at all.
What is the best model for Hermes Agent?
For most people, GLM-5.2. It posts a composite coding score of 68.8, level with Google’s Gemini 3.1 Pro, at a blended cost near $2.15 per million tokens against Gemini’s $4.50. If you need the absolute strongest reasoning and can pay for it, use Claude Opus 4.8. If the bill is what keeps you up at night, use DeepSeek V4. Everything below is how those three shake out, and when a fourth choice beats all of them.
One thing to settle first, because it decides which models are even eligible. Nous states plainly that Hermes needs at least 64,000 tokens of context for agent use with tools, because the system prompt, tool schemas, and running conversation state have to fit with room to work. Run Ollama locally and you have to set OLLAMA_CONTEXT_LENGTH=64000 by hand or the agent quietly falls apart mid-task. 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 Fable 5 | $20 | 76.5 |
| Claude Opus 4.8 | $10 | 74.3 |
| Gemini 3.1 Pro | $4.5 | 68.8 |
| GLM-5.2 | $2.15 | 68.8 |
| Claude Sonnet 5 | $6 | 66.4 |
| Qwen3.7 Max | $3.75 | 66 |
| Kimi K2.7 Code | $1.71 | 60.8 |
GLM-5.2 is the point that breaks the pattern. Same capability as Gemini 3.1 Pro, less than half the cost, and Z.ai released it as open weights under the MIT license, so the same model you rent through a gateway is the one you can run on your own hardware later. Opus 4.8 and Fable 5 own the top of the chart, but you pay steeply for those last few points. And Kimi K2.7 Code sits at the bottom-left: cheapest of the plotted set, still a real coding model.
Best raw performance: Claude Opus 4.8
When a task is genuinely hard, a long refactor across a repo, 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 4.8 posts a 74.3 composite coding score and holds tool chains together over long horizons better than anything in the mid-tier. It runs $5 per million input tokens and $25 output, roughly $10 blended, which is real money at volume but trivial next to an hour of debugging a hallucinated fix.
Claude Fable 5 scores higher, 76.5, and if you want the top of the chart it is the one. The catch is access. Fable 5 was caught in a June 2026 US export-control directive that suspended it for a stretch; Anthropic began restoring general access in early July 2026 on a staged rollout. Availability may still be uneven depending on who and where you are, so treat it as the ceiling with an asterisk, not the safe default.
Best value: GLM-5.2
Here is the pick I would hand almost anyone starting out.
GLM-5.2 from Z.ai matches Gemini 3.1 Pro’s coding score exactly, 68.8 to 68.8, while costing $1.40 input and $4.40 output against Gemini’s $2 and $12. On the blended number that is $2.15 versus $4.50. You are getting Gemini-Pro-class coding for the price of a budget model, and Artificial Analysis ranked GLM-5.2 the strongest open-weights model available, behind only the closed frontier labs. It comfortably clears the 64K context rule, and because it is an open-weights release you are not locked to one vendor’s endpoint: rent it now through a gateway, move it in-house when your volume justifies a GPU.
The trade is that it is not quite frontier. On the hardest problems Opus still pulls ahead. But for the daily mix of coding, tool use, and file wrangling that a personal agent actually does, GLM-5.2 is the best money-to-capability ratio on the board.
Cheapest that still works: DeepSeek V4
If the goal is to run an agent as close to free as an API allows, DeepSeek V4 is the answer, and it is not close.
At $0.435 per million input tokens and $0.87 output, it lands near $0.54 blended, an order of magnitude under the frontier. It ships a 1M-token context window, so the 64K floor is a non-issue, and it supports tool calls, which is the actual make-or-break for agent work. DeepSeek does not publish a directly comparable composite coding score, so I would not put it on the hardest tasks, but for high-volume, well-defined chores, summarizing, tagging, routine edits, drafting, it is the model that lets you stop watching the meter.
| Item | Value |
|---|---|
| Claude Opus 4.8 | $10.00 |
| Claude Sonnet 5 | $6.00 |
| Gemini 3.1 Pro | $4.50 |
| Qwen3.7 Max | $3.75 |
| GLM-5.2 | $2.15 |
| Kimi K2.7 Code | $1.71 |
| DeepSeek V4 | $0.54 |
Best local and private: open weights through Ollama
Some work should never leave your machine. Client code under NDA, anything touching personal data, or just a preference not to stream your keystrokes to a third party.
Hermes supports fully local inference through Ollama, vLLM, SGLang, llama.cpp, and LM Studio, and here the open-weights models earn their keep. GLM-5.2 is again the standout because you can rent it and self-host the identical model. Qwen’s open releases are the other strong option, with smaller variants that fit consumer GPUs.
Be honest with yourself about the hardware, though. Running a frontier-class open model at a real 64K context is a datacenter-GPU job, not a laptop one. On a single consumer card you will be running a smaller model and accepting a capability drop, which is often a fine trade for the privacy but is a trade all the same. The free path is real, but “free” here means you already own the silicon.
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 |
|---|---|---|
| Hardest coding and long refactors | Claude Opus 4.8 | Highest reliable coding score, best long-horizon tool use |
| Best all-round daily driver | GLM-5.2 | Gemini-Pro-class coding at roughly half the cost, open weights |
| High-volume cheap chores | DeepSeek V4 | An order of magnitude cheaper, clears 64K, supports tools |
| Long context and big-memory recall | DeepSeek V4 or Gemini 3.1 Pro | 1M-token windows for large tool outputs and history |
| Fully local and private | GLM-5.2 or Qwen (open weights) | Nothing leaves your box; needs a capable GPU |
| Cheapest frontier-lab option | Gemini 3 Flash | $0.50 input, tool support, from a major lab |
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-5.2. If a task proves too hard, escalate that job to Opus for a day. When DeepSeek ships a cheaper tier, move your 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 will turn 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?
- For most users, GLM-5.2: it matches Gemini 3.1 Pro on composite coding score while costing roughly half as much per token, and its open weights let you self-host later. Use Claude Opus 4.8 when a task is hard enough that a wrong answer costs more than the tokens, and DeepSeek V4 when you want the lowest possible bill. All three clear the 64,000-token context minimum Hermes needs for tool use.
- What is the cheapest model for Hermes Agent?
- DeepSeek V4, at roughly $0.435 per million input tokens and $0.87 output, which lands near $0.54 blended, about a tenth of frontier pricing. It ships a 1M-token context window and supports tool calls, so it clears Hermes requirements comfortably. It is best for high-volume, well-defined chores rather than the hardest reasoning tasks.
- Can Hermes Agent run local models for free?
- Yes. Hermes supports fully local inference through Ollama, vLLM, SGLang, llama.cpp, and LM Studio, with no API key. Open-weights models like GLM-5.2 and Qwen are the strongest local picks. The catch is hardware: running a frontier-class open model at a real 64K context needs a capable GPU, so "free" assumes you already own the machine.
- What context window does Hermes Agent need?
- At least 64,000 tokens, per Nous documentation, because the system prompt, tool schemas, and conversation state have to fit with room for multi-step work. On Ollama you must set OLLAMA_CONTEXT_LENGTH=64000 explicitly. A model that only offers a small context will fail on agent tasks regardless of how well it benchmarks elsewhere.
- Does the model even matter if the harness does the work?
- It sets the ceiling. Harness engineering can move a fixed model double digits on a benchmark, so the software around the model matters enormously, but a stronger model raises the ceiling that software works against. The practical move is to pick a good-value model like GLM-5.2, invest in the harness, and swap the model when the economics or the task demand it.
Sources
- Nous Research (2026). Hermes Agent documentation: AI providers, model catalog. https://hermes-agent.nousresearch.com/docs/integrations/providers
- Anthropic (2026). Claude pricing and Claude Fable 5 / Mythos 5 announcement. https://claude.com/pricing
- DeepSeek (2026). API pricing (official documentation). https://api-docs.deepseek.com/quick_start/pricing
- Artificial Analysis (2026). Model comparison: coding and intelligence indices. https://artificialanalysis.ai/models
- Capital & Compute. AI model registry (per-token rates, verified July 2026). https://capitalandcompute.net/ai-models/
Prices and capability scores were verified against the sources above on July 5, 2026. Per-token rates and benchmark composites change frequently; the value picks reflect the market as of that date. See our editorial standards.