Cohere adds Command A Reasoning, a 111B-parameter model with 256K‑token: what developers gain

Cohere on 2025 — 08-21 released Command A Reasoning, identified as command — a-reasoning-08-2025, a 111 billion‑parameter model designed for agentic applications where models must plan, call external tools, and carry out multi‑step problem solving rather than only return final answers. The company emphasizes the model’s ability to emit intermediate chains of thought and to act autonomously using external resources, positioning it for workflows that require multi‑step coordination and transparent reasoning.

Technically, the model supports an extended 256K token context window and can produce outputs up to 32K tokens, enabling longer state, extended memory and multi‑document responses. Developers can access the model through Cohere’s Chat API endpoint; the release follows the provider’s existing chat patterns so teams familiar with the API can integrate without changing core request structures.

Command A Reasoning exposes explicit intermediate reasoning in the API stream: responses can include content items labeled with content.type values such as "thinking" for internal reasoning and "text" for final user‑facing output. Client code can therefore present, log or suppress intermediate chains of thought independently from the final answer. The API also supports streaming and non‑streaming usage, and documentation includes a non‑streaming example that iterates response.message.content items to separate thinking and text content.

The model is framed as the strongest tool user in Cohere’s Command family, with built‑in behavior for proactive resource use and autonomous action. It is multilingual — supporting English plus 22 other languages — which Cohere highlights for businesses that need reasoning and agentic capabilities outside English. The combination of longer context, higher output limits and enhanced tool use aims to reduce friction when chaining external APIs or utilities for complex tasks.

Cohere provides customization controls for how the model surfaces internal reasoning. Developers can toggle explicit thinking behaviour with a thinking parameter and manage intermediate token consumption with a configurable thinking budget that caps token usage devoted to internal reasoning. Release notes direct builders to dedicated Reasoning documentation for details on token budget mechanics, advanced configuration and recommended practices for steering agentic behaviors. Taken together, the release signals a step toward production‑grade agentic assistants and complex automation: the model’s larger context window and higher output ceilings support extended interactions and long‑form plans, while explicit thinking outputs and tool‑use capabilities make it easier to compose and audit multi‑step agent workflows.