GPT-5.6 Release Notes: Advancing Agent Price-Performance

OpenAI has introduced the GPT-5.6 model family, designed to make frontier-level agent performance more affordable while expanding the capabilities of autonomous agents. The release focuses on increasing price-performance efficiency, allowing developers to maintain high accuracy while reducing token expenditure and operational costs.

Enhanced Price-Performance and Model Selection

GPT-5.6 provides higher accuracy at lower reasoning efforts compared to previous generations. In testing on the Agents’ Last Exam benchmark, GPT-5.6 Sol at "low" reasoning effort outperformed GPT-5.5 at "high" reasoning effort when using a constant harness.

Developers can now achieve flagship-level performance using more cost-optimized models within the 5.6 family, such as Luna and Terra, by utilizing increased test-time compute. This shift is evidenced by the BrowseComp search-based benchmark:

GPT-5.5 (Extra High): Scored 84.36% at a cost of $33.27.
GPT-5.6 Luna (Extra High): Scored 84.04% at a cost of $1.33.

Smaller models like Luna and Terra are specifically recommended for high-volume workloads, latency-sensitive interactions, and repetitive steps in agentic workflows, such as data extraction from handwritten memos.

Responses API Updates for Agent Architecture

OpenAI has updated the Responses API with three architectural interventions to improve agent efficiency:

1. Reasoning Persistence and Compaction

GPT-5.6 allows reasoning to be persisted across model turns and utilizes native compaction to compress long-running conversations. These features prevent the model from losing coherence or needing to reconstruct context over long task horizons. On the ARC-AGI-3 benchmark, enabling retained reasoning and compaction increased GPT-5.6 Sol's score from 13.3% to 38.3% while using approximately 6 fewer output tokens.

2. Native Multi-Agent Orchestration

The Responses API now supports native multi-agent orchestration, enabling a primary agent to delegate tasks to subagents who work in parallel. This distribution of reasoning across parallel workstreams increases task completion speed and overall intelligence. This orchestration is also the foundation for the "ultra" capability setting in ChatGPT.

3. Programmatic Tool Calling

To reduce "context rot" and lower costs, programmatic tool calling allows GPT-5.6 to write JavaScript to orchestrate tools and process outputs (such as filtering or aggregating data) outside the model's context window. This reserves model tokens for tasks requiring judgment rather than deterministic data manipulation.

Prompt Caching Improvements

OpenAI has extended the prompt cache Time-to-Live (TTL) to a minimum of 30 minutes. Additionally, developers can now set cache breakpoints deterministically within a model's context window to improve cache hit rates. The use of a prompt_cache_key is recommended to increase the likelihood of requests landing on the same inference engine, thereby reducing latency.