OpenAI outlines GPT-5.6 tools for lower-cost production agents

OpenAI has published a builder guide for GPT-5.6, presenting the model family and new Responses API controls as a way to reduce the cost of production agents without sacrificing longer-horizon performance. The company highlights GPT-5.6 Sol, Luna and Terra, alongside retained reasoning, native compaction, multi-agent orchestration, Programmatic Tool Calling and expanded prompt caching.

One cited result comes from ARC-AGI-3: GPT-5.6 Sol scored 13.3% with a standard harness. When retained reasoning and compaction were enabled, OpenAI reports a score of 38.3% with roughly six times fewer output tokens. OpenAI frames that change as an architectural gain rather than a modification to the model itself.

Model selection shifts toward smaller models

OpenAI says the 5.6 family changes the trade-off between flagship models and cheaper variants for long-context and tool-calling work. With more test-time compute, Luna and Terra can often perform similarly to GPT-5.4 and GPT-5.5 while costing substantially less, the company says.

On the BrowseComp search benchmark, OpenAI compares GPT-5.5 at Extra High reasoning, which scored 84.36% at a total cost of $33.27 three months earlier, with GPT-5.6 Luna at Extra High reasoning. Luna scored 84.04% for $1.33 at launch. The guide positions the smaller models for high-volume operations, latency-sensitive exchanges and recurring stages in agent workflows.

The pricing and product context includes OpenAI pricing changes for Luna and Terra as OpenAI positions Luna and Terra for lower-cost use cases, while Sol remains the higher-capability option in the family. For example, the guide suggests using Terra or Luna to extract information from handwritten legal memos before a later agentic analysis step, rather than applying a frontier model across the entire workflow.

Responses API controls move work out of the context window

The updated Responses API is designed around three interventions. Retained reasoning and native compaction allow a model to preserve work across turns and compress long-running conversations. Native multi-agent orchestration lets a primary agent delegate parallel workstreams to subagents and synthesize their results. Programmatic Tool Calling lets GPT-5.6 write JavaScript to coordinate tools, execute independent calls in parallel and process outputs outside the model context.

That last capability is aimed at workflows that combine judgement with deterministic operations. If an agent retrieves filings, filters them by date and combines their contents, the filtering and aggregation can happen in code while the model focuses on interpreting relevant transactions. Rogo reported that GPT-5.6 using Programmatic Tool Calling matched its rubric quality in financial-research evaluations while using 21% fewer input tokens.

Cache controls support repeated agent work

Across the family, OpenAI has extended the minimum prompt-cache time-to-live to 30 minutes and added deterministic cache breakpoints within a context window. It also recommends an appropriate prompt_cache_key to increase the chance that requests with the same prefix reach the same inference engine, reducing latency.

For businesses building agents, the implication is to evaluate the workflow rather than defaulting every stage to the most capable model. Teams can assign extraction and repeated operations to smaller models, retain reasoning only where continuity matters, move deterministic tool processing into code, and compare quality, token use and response time under their own production harnesses.