OpenAI Extends Reasoning Models with Multi-Turn State Retention

Summary

OpenAI has released an update to its reasoning-model API capabilities, allowing models to maintain internal reasoning state across multiple conversation turns rather than discarding it after each response. This enhancement applies to OpenAI's latest reasoning models, including GPT-5.5 and GPT-5.6, which use internal reasoning tokens to plan, use tools effectively, inspect alternatives, and solve complex multi-step tasks. The updated functionality is accessed through OpenAI's Responses API, which offers improved model intelligence and performance compared to the legacy Chat Completions API.

The update addresses a key limitation in reasoning-model deployment for complex, iterative workflows. By preserving reasoning context across turns, developers can now build more sophisticated agentic systems and multi-step problem-solving applications with reduced redundancy and improved efficiency. The reasoning models support configurable effort levels (from minimal to xhigh), allowing developers to balance quality and latency based on their use case. GPT-5.6 also introduces a pro reasoning mode for even more challenging tasks.

OpenAI is offering multiple models at different capability and cost tiers: GPT-5.6 and GPT-5.5-pro for the highest intelligence, GPT-5.5 for balanced performance, and GPT-5.4/GPT-5.4-mini for lower cost and latency. The feature is particularly beneficial for complex problem-solving, scientific reasoning, coding tasks, and agentic workflows—areas where iterative reasoning and tool use are critical.

• Enhanced capabilities for multi-step agentic workflows, scientific reasoning, and complex problem-solving tasks