For Hermes Agent in September 2026, Claude Sonnet 5 is a practical starting point for daily automation, Claude Opus 5 or GPT-5.6 Sol fits harder work, and DeepSeek V4.1 Flash or GLM-5.3 Flash keeps routine runs inexpensive. These are starting points, not a controlled Hermes benchmark. A model that looks strong on a coding leaderboard can still fail a long tool loop or cost more through retries.
Hermes lets you change the main model without changing the agent. Its model configuration guide also separates the main model from auxiliary models used for smaller jobs. That means you can spend on the reasoning loop and choose cheaper models for compression or summarization. The model IDs below were checked against haimaker’s live catalog on September 16, 2026; availability and pricing can change.
Which model should you start with?
Start with the task your agent performs most often. In recent Hermes user reports, people use different models for daily work, hard reasoning, and local private tasks. The table gives a short list to test on your own prompts, tools, and failure cases.
| Hermes workload | Model to test first | haimaker model ID | Tradeoff |
|---|---|---|---|
| Daily assistant and automation | Claude Sonnet 5 | anthropic/claude-sonnet-5 | Reliable premium default; costs more than budget models |
| Production coding and difficult debugging | Claude Opus 5 | anthropic/claude-opus-5 | Stronger reasoning ceiling; higher cost per run |
| Hard planning with an OpenAI model | GPT-5.6 Sol | openai/gpt-5.6-sol | Another premium option; compare token use on long runs |
| Long-context research | Gemini 3.8 Flash | google/gemini-3.8-flash | Large input window; check answer quality on your documents |
| Open-weight model through a cloud API | Kimi K3 | moonshotai/kimi-k3 | Flexible alternative; test provider latency and tool use |
| Budget background work | DeepSeek V4.1 Flash or GLM-5.3 Flash | deepseek/deepseek-v4.1-flash or z-ai/glm-5.3-flash | Lower token cost; retries can erase the savings |
These choices reflect current availability and reported use, not a claim that every model has passed the same Hermes test suite. The Hermes community’s August provider roundup also places Sonnet and Opus in the quality tier while budget users favor DeepSeek and GLM. Prices in that roundup are a snapshot, so check the live catalog before setting a budget.
How do you switch models in Hermes Agent?
Run hermes model, choose Custom endpoint, and supply the base URL, API key, and model ID. Hermes keeps the endpoint choice in its configuration. You can then compare models on the same task instead of inferring their behavior from unrelated benchmarks.
export HAIMAKER_API_KEY=your-haimaker-key
npx -y @haimaker/connect --hermes
@haimaker/connect writes the Hermes endpoint configuration and verifies the key. For manual setup, use https://api.haimaker.ai/v1 as the base URL and copy a current ID from the model catalog. The Hermes custom provider guide covers per-provider URLs and timeout fixes.
Changing the model is only one variable. If a research run fails, check search, extraction, browser access, and tool permissions before blaming the model. Hermes users have reported that those dependencies can dominate the result.
Which model fits each job?
Daily automation: Claude Sonnet 5 is the first cloud model to try when Hermes handles messages, files, and scheduled tasks. Run a representative multi-step task and count failed tool calls, retries, and total tokens. If the task is routine, compare it with a budget model before committing to a premium default.
High-stakes coding: Claude Opus 5 and GPT-5.6 Sol are candidates when the agent edits a production repository or must recover from difficult failures. Compare them on the same issue with the same tools. A single successful run is useful evidence, but repeated runs reveal whether the model handles interruptions and mistakes consistently.
Long documents and research: Gemini 3.8 Flash has a large input window in the current haimaker catalog. That makes it worth testing when Hermes needs to carry many documents or tool results. A large window does not guarantee accurate recall; verify the answer against the source files.
Always-on background work: DeepSeek V4.1 Flash and GLM-5.3 Flash are low-cost options. MiniMax M3 (minimax/minimax-m3) remains available if you already use it. Give each candidate a bounded task, a retry limit, and a spending cap. A cheap model that repeats failed actions can cost more than a stronger model that finishes once. Keep a premium model available for escalation rather than sending every task to it.
Use one haimaker.ai key to compare these models on the same Hermes task. Keep the model that completes the work reliably at an acceptable total cost.
TEST MODELS WITH ONE KEYShould you run Hermes on a local model?
Hermes can use a local OpenAI-compatible endpoint. Its local-model guide recommends choosing a build that fits your actual memory budget, including context state and runtime buffers. A model that barely loads may have too little room left for the context a long agent run needs.
Local use makes sense when private data must stay on your machine or you already have the hardware. The September Hermes model discussion includes users pairing a local main model with cloud models for harder tasks. Test tool calls and context handling before leaving a local model unattended. Our Ollama coding models guide covers hardware tiers, and cheapest APIs for coding agents covers the cloud side.
How should you compare the finalists?
Run the same three tasks on each candidate: a routine automation, a long tool sequence, and a failure that requires recovery. Record completion, incorrect tool calls, retries, elapsed time, and total cost. Keep the task instructions and tool access fixed so the model is the main variable. If a model wins only on a benchmark but fails your workflow, choose the one that finishes your work.
Hermes changes quickly, and model catalogs move with it. Recheck model IDs in the live catalog when you configure the agent, then repeat this small comparison when a new model looks promising.
Frequently asked questions
What is the best model for Hermes Agent in 2026?
Claude Sonnet 5 is a strong starting point for daily Hermes automation. Move to Claude Opus 5 or GPT-5.6 Sol when a difficult task justifies the premium, and test a budget model such as DeepSeek V4.1 Flash or GLM-5.3 Flash for routine background work. Your own tool-heavy tasks are the final test.
What is the cheapest model that works well in Hermes Agent?
DeepSeek V4.1 Flash and GLM-5.3 Flash are current low-cost cloud options in the haimaker catalog. A local model avoids per-token API charges but needs suitable hardware and can struggle with long tool loops. Test reliability and total task cost before choosing a default.
How do I change the model Hermes Agent uses?
Run hermes model, choose Custom endpoint, and enter the base URL, API key, and model ID. With haimaker.ai, use https://api.haimaker.ai/v1 and a model ID from the live catalog. You can also run npx -y @haimaker/connect --hermes to configure the endpoint.
Comparing agents rather than models? See Hermes vs Codex. Setting up the endpoint itself? See the Hermes custom provider guide.