- Published on
Llm 0.36
- Authors
- Name
- DeepHorus
The LLM command-line tool was released in version 0.36 on 22nd September 2026, adding support for OpenAI's newest GPT-6 models alongside several plugin-facing improvements. The update registers two model identifiers, gpt-6-sol for GPT-6 Sol and gpt-6-luna for GPT-6 Luna, making both immediately available to users of the tool. The release lands amid a notably busy week for the industry, with rival announcements arriving from Anthropic and independent labs. It also coincides with fresh coverage of a widening price war across frontier model providers.
Beyond the new model entries, the headline change concerns how plugins describe model capabilities. Model plugins can now declare supports_conversation = False for models that accept only single-turn prompts, a flag that formalises a limitation previously handled informally. When such a model receives assistant or tool history, LLM raises llm.ConversationNotSupported rather than forwarding a malformed request. In addition, the llm chat command rejects these models before a session ever starts, preventing wasted calls and confusing mid-session failures.
The first plugin to adopt the new declaration is llm-typesafe, which targets strongly typed interactions with models. Tracked as issue #1702, the change gives plugin authors a standard way to signal that a model cannot sustain multi-turn dialogue. Documentation has been updated under a new section titled Models that do not support conversations, outlining expected behaviour for developers and users. The move should reduce trial-and-error when connecting custom plugins to newly released endpoints.
Two smaller but user-facing refinements shipped in the same version. Under issue #1692, reasoning traces shown in the Markdown output of llm logs are now wrapped in
tags, allowing long chains of thought to be collapsed by default. Issue #1701 covers related log-rendering work included in the release. Together, these tweaks keep verbose output readable without hiding it entirely from users auditing past sessions.
The timing matters because the surrounding ecosystem is moving fast. On the same day, coverage highlighted Claude Opus 5.5 alongside GPT-6 Sol and GPT-6 Luna, framing the launches as the start of a new price war. Just a day earlier, Jev unveiled what it called System One, or Decision Models, described as a new shape of LLM architecture. Recent weeks have also seen practical demonstrations, such as generating running routes with GPT-6 Astra and ChatGPT Work on 12th September 2026.
For LLM users, version 0.36 offers both immediate access to OpenAI's latest models and clearer safeguards when those models hit their limits. The capability flag establishes a shared convention that other plugins are likely to follow as single-turn or reasoning-focused models proliferate. Looking ahead, further releases will probably extend the flag to other constraints, such as context-window or modality limits. Developers can track progress through the project's ongoing issue tracker as the GPT-6 family expands.
