Hermes Agent: a new model in the picker does not grant every account access
Claude 5.5, GPT-6 and Gemini 4 Argon appear in the 0.22 notes. Check the provider and account behind a model before changing your setup.

Generated image · SecuFocus
At a glance
Key points
The 0.22 cycle expands Hermes’s model choices and makes some subscription limits more visible. A name in the picker is not enough, though: availability, billing and context size also depend on the provider, account and transport used.
A catalog entry is not a universal entitlement
The 0.22.0 notes name Claude Opus, Sonnet and Haiku 5.5, GPT-6 and GPT-6.1 Sol, and Gemini 4 Argon on routes where offered. These are additions to Hermes, not proof that every account has received every model.
The documentation distinguishes subscription routes, such as ChatGPT/Codex through OAuth, from API-key access. Two routes to similarly named models can have different quotas and billing. Refresh the model list if needed, then check the selected account and provider before starting work.
Speed, reasoning and context are separate controls
The 0.22 cycle adds reasoning-effort selection to model pickers. The documentation treats Fast and Ultrafast as separate service tiers: an option appears only where offered for that model and account. An accepted request does not guarantee the provider served the requested tier.
The GPT-6.1 Sol-900k suffix is documented as a local context-budget choice in Hermes, removed from the model ID sent to OpenAI. The account catalog still limits the budget, and the backend may reserve part of that window. Do not treat the suffix as a promise that every conversation will fit without compaction.
Favorites help selection; a fallback changes the destination
Desktop can lift favorite models to the top of its picker. The 0.22.0 notes also announce remaining-capacity and rate-limit-reset indicators for supported subscriptions. These indicators are not a universal billing statement.
Choose a fallback provider with the same care as the primary one. Fallback can send a request through another route when the first fails. Before configuring it, check whether the destination is appropriate for the project’s data and whether the cost is acceptable. A local primary model does not make the fallback local.
The release also describes visible, bounded recovery cycles for some transient errors. An announced wait is not necessarily a stuck agent, and a configured fallback does not prove a reply still came from the primary provider.
Avoid repeatedly switching models in a long chat
The Desktop documentation warns that switching models mid-chat rebuilds the prompt cache for that model: the next turn can reread the entire history at the applicable input rate. For a lasting change, a fresh conversation may be preferable to repeated switching.
Start with a short task containing no sensitive data, then check the reported model and route. Auxiliary models and scheduled jobs can follow other settings; selecting a model for the current chat does not reconfigure them all. SecuFocus has run neither a quality comparison nor a cost measurement for this article.
Check and explore
Sources for this article
Numbers connect each reference to the passages that use it. Dates show when the documentation was consulted.
- Documentation Hermes : fournisseurs et modes d’authentification ↗hermes-agent.nousresearch.com ·
- Documentation Hermes : fournisseurs de secours ↗hermes-agent.nousresearch.com ·
This article draws on the sources above. The exercises are for you to try on your devices; SecuFocus does not present them as tests carried out by its editorial team. Interfaces and features can change. Method and corrections.
Cite this article
Keep this reference with the article when you save or share it.