OpenAI’s GPT-5.6 guide focuses on cheaper, more efficient agents

OpenAI’s new builder’s guide for GPT-5.6 outlines how startups are using the model family and updated Responses API features to build more capable agents at lower cost.

Smaller models, broader workloads

The guide argues that model selection is becoming more nuanced than simply choosing the flagship model at maximum reasoning effort. OpenAI says GPT-5.6 Luna and Terra can approach the performance of earlier frontier models on some long-horizon tasks while costing considerably less.

The examples are significant for teams operating agents at scale. Hypha reported that Luna retained 98% of GPT-5.5’s document-extraction accuracy at one-eighteenth the cost. Browser Use said Luna completed 78% of 106 difficult browser tasks for about $14, compared with roughly $235 for its leading model. OpenAI also cites a BrowseComp comparison in which Luna delivered nearly the same score as GPT-5.5 at a much lower cost.

API controls become part of the architecture

The guide highlights three ways developers can reduce waste in agentic systems: preserve reasoning between model turns, use native compaction to manage long conversations, and coordinate parallel work with multi-agent orchestration.

Programmatic Tool Calling is another major focus. It allows the model to write JavaScript that runs tools, processes results, and handles deterministic filtering or aggregation outside the context window. That leaves the model to make decisions instead of spending tokens moving data around.

OpenAI reports that combining retained reasoning and compaction raised GPT-5.6 Sol’s ARC-AGI-3 score from 13.3% to 38.3% while using about six times fewer output tokens. For AI builders, the message is clear: better agent economics may come as much from workflow design and API primitives as from upgrading the model itself.

Source: OpenAI News