Best AI model for OpenClaw and Hermes: which to use for what

Use a cheap, fast model for most of what an AI agent does: email, admin, quick lookups, summaries and scheduled jobs. In Heyron that's Default. Outside Heyron, DeepSeek V4.1 Flash is the most-used model in Hermes Agent and the second most-used in OpenClaw.

Step up to a mid-priced model, such as Claude Sonnet 5.5, GPT-6.1 Sol or Gemini 4 Argon, for writing people will read, research you'll act on, and code. Save Claude Opus 5.5, which leads the broadest independent rankings, for the few jobs where a wrong answer costs you more than the model does. A mid-priced model costs about 20 times as much per request as a cheap one, and Opus twice that again, so switch back when the hard part is done.

This guide covers which model to use for each job, what each one costs per request, what OpenClaw and Hermes users actually run, and how to switch models in Heyron. Prices were checked on November 5, 2026, from OpenRouter's public price list, and the test results are from the same week. Both change every month.

Which model should your agent use for each job?

Why does the model matter more for an agent than for a chatbot?

When you chat with ChatGPT, each message is short. An agent's requests are much bigger. Every time it answers or takes a step, it sends the model its instructions, the tools it can use, what it remembers about you, and the conversation so far. Across Heyron agents in October 2026, the typical request carried about 43,000 tokens, roughly 30,000 words, to get back a reply of a few hundred words.

Two more things multiply the bill:

• It works when you don't. A scheduled job that runs every 15 minutes runs about 2,900 times a month, and each run is at least one request. Messages from Telegram or WhatsApp and checks that run in the background all count too.
• Thinking costs extra. Higher effort levels make the model think before it answers, and those thinking tokens are billed at the output price, which is five times the input price on most models.

So a model that costs a few cents more per message in a chat window can cost dollars more a day once it's doing an agent's work.

What does each model cost per request?

For a sense of a whole job: a research task from Heyron's Roadmap took our own agent about two and a half minutes and cost 2.2¢ on Default. The same task would cost about 90¢ on Claude Sonnet 5.5, $1.70 on Claude Opus 5.5 and $4.30 on Claude Fable 5.1.

Real costs can be lower. Most models give a discount when the agent re-sends the same context within a few minutes, which is called prompt caching. On Claude Opus 5.5, cached input costs 5% of the normal price. But OpenClaw's docs note that changing the model or the effort level mid-chat resets the cache, one more reason to switch models in a new chat.

What do OpenClaw and Hermes users actually run?

Not the models at the top of the leaderboards. On OpenRouter in early November 2026, the most-used models in OpenClaw were Step 5 Preview, DeepSeek V4.1 Flash, MiniMax M3.1 Flash, GLM 5.3 Flash and a newer DeepSeek V4.1 Flash snapshot. In Hermes Agent, DeepSeek V4.1 Flash came first. Claude Opus 5.5 was the only premium model in either app's top ten.

That isn't people being cheap for the sake of it. Low-cost models have caught up on the routine work agents do most. On τ²-bench, a test of using tools while following a set of rules, DeepSeek V4.1 Flash scores 75%, level with Claude Opus 5.5, and Gemini 3.5 Flash-Lite ranks ninth of 142 models. On AutomationBench, MiniMax M3.1 Flash beats Opus outright.

The gap opens up on long, open-ended work. On GDPval, Artificial Analysis's test of real office tasks, Opus 5.5 comes first and DeepSeek V4.1 Flash is well behind. That's why hard code, research across many sources and decisions with lots of trade-offs are still worth a stronger model.

What's the cheapest model that works well with OpenClaw?

DeepSeek V4.1 Flash, at about 0.2¢ per typical request. It's the most-used model in Hermes Agent, second in OpenClaw, and it scores level with Claude Opus 5.5 on τ²-bench. Close behind:

• Claude Haiku 5.5 and GPT-6 Luna, at about 0.45¢. Anthropic recommends Haiku for sorting, pulling out details and routing, and Luna is one of the most-used models on OpenRouter. Luna is weak at code.
• MiniMax M3.1 Flash, at about 0.7¢, which scores above Claude Opus 5.5 on AutomationBench.

Two cautions. OpenClaw's docs advise against older, weaker models for agents that use tools. And for open models, the cheapest price on OpenRouter often comes from a host running a compressed version, which can make more mistakes. In Heyron, Default takes care of both: it's a fast, low-cost model chosen for you and kept current.

Which model is best for Hermes Agent?

The same choices apply, with one difference: Hermes sends bigger requests. Across Heyron agents in October, the typical Hermes request carried about 70,000 tokens, against 38,000 on OpenClaw. So every model costs nearly twice as much per request on Hermes, and a cheap everyday model matters even more.

DeepSeek V4.1 Flash is the most-used model in Hermes on OpenRouter. For harder work, Claude Sonnet 5.5 is a good step up; OpenRouter's own Hermes tutorial uses Claude Sonnet as its default. Hermes needs a model that can read at least 64,000 tokens at once, which every model in this guide can.

When should you turn up the effort?

The effort picker sets how long the model thinks before it answers. Thinking is billed as output, so it adds up fast. Artificial Analysis measured Claude Sonnet 5.5 at about $1.64 per task on high effort and $10.02 on max: six times as much for the same work.

Leave effort at its normal setting for routine work. Turn it up for one hard step, like a tricky bug or a plan with many trade-offs, then turn it back down. If a cheap model still gets a job wrong at high effort, try a stronger model at normal effort rather than pushing effort to the top.

How do you switch models in Heyron?

In any chat on your dashboard, the model picker sits under the message box. Default is at the top; type to search the rest of the list. The effort picker next to it sets how long the model thinks, from None up to Max or Ultra. On OpenClaw, the model you pick stays with that chat.

Everything that runs without you, like scheduled jobs, background checks and messages from Telegram, WhatsApp, Discord or Slack, uses your agent's default model, which is Default unless you change it. If you want a different default, open a support ticket and we'll switch it for you.

The habit that saves the most: start a new chat on a stronger model for the hard part, then go back to a chat on Default for the follow-up work. See long chats and starting fresh and AI usage, models and top-ups.

How do you test a model on your own work?

Leaderboards tell you what a model can do on someone else's tasks. A ten-minute test on your own tells you more.

Mistakes to avoid

• Leaving a premium model on. You switched to Opus to untangle one problem, then spent the afternoon on routine edits. Switch back when the hard part's done.
• Putting scheduled jobs on an expensive default. A job that runs every hour runs about 720 times a month. On Claude Opus 5.5 that's at least $130.
• Choosing from chat leaderboards alone. A model that writes a lovely essay can still fumble an agent's tool calls. Test it on a real job first.
• Paying for thinking you don't need. High effort helps on hard problems. On routine work it mostly adds cost and waiting time.
• One endless chat. Every message re-sends the conversation so far, so a chat that's run for weeks costs more per message on any model.

Try it on an agent of your own

Heyron gives you an AI agent of your own on OpenClaw or Hermes, running on its own computer, with hundreds of models in one picker and AI usage included every month. You can switch models per chat without opening an account with each provider. Members also get the Solo Software Studio community, where you can ask people building the same kind of one-person business what's working for them. See what's included.

Frequently asked questions

What's the best AI model for OpenClaw?

For most jobs, a low-cost model like DeepSeek V4.1 Flash, or Default in Heyron, is good enough and costs about 0.2¢ a request. For writing, research and code you'll rely on, step up to Claude Sonnet 5.5 or GPT-6.1 Sol. Claude Opus 5.5 leads the broadest independent rankings, but a request costs about 80 times as much as on Default.

What's the cheapest model that still works for an agent?

DeepSeek V4.1 Flash, at about 0.2¢ per typical request. It's the most-used model in Hermes Agent and scores level with Claude Opus 5.5 on τ²-bench, a tool-use test. Claude Haiku 5.5 and GPT-6 Luna cost about twice as much and are also good picks for high-volume work.

How much does it cost to use OpenClaw?

The software is free. Run it yourself and you pay for a server, about $5–20 a month, plus AI usage, which ranges from about 0.2¢ to 45¢ per request depending on the model. See self-hosting OpenClaw vs Heyron for the full comparison. Heyron is $49 a month, including the agent's computer and $15 of AI usage.

What's the best model for Hermes Agent?

The same models as for OpenClaw: a low-cost model for everyday work, and Claude Sonnet 5.5 or GPT-6.1 Sol for harder jobs. Hermes sends larger requests than OpenClaw, about 70,000 tokens against 38,000 in our data, so the model you choose affects cost even more.

Is Claude or GPT better for an AI agent?

It depends on the job. Claude Opus 5.5 leads the broad rankings and the GDPval office-work test, and Claude Sonnet 5.5 is strong at code. GPT-6.1 Sol gets close to the top models for less, and GPT-6.1 Terra is good value for writing. Gemini 4 Argon tops a test of email and app tasks.

Does a more expensive model always do better?

No. Claude Fable 5.1 costs 2.5 times as much as Claude Opus 5.5 and scores lower on most agent tests, and MiniMax M3.1 Flash beats Opus on a test of email, Slack and CRM tasks. Pick by the job, and test on your own work.

How do I change the model in Heyron?

Use the model picker under the message box in any chat. Default is at the top. Scheduled jobs and messaging apps use your agent's default model; open a support ticket to change it.