Rethinking Model Selection and Reasoning Effort

GPT-5.6 introduces a significant shift in price-performance, allowing developers to move away from the default "frontier model for everything" approach. The new model family—specifically the Luna and Terra variants—can match the performance of previous flagship models (like GPT-5.4 and 5.5) at a fraction of the cost. For instance, GPT-5.6 Luna achieved an 84.04% score on the BrowseComp benchmark for $1.33, compared to the $33.27 cost of GPT-5.5 for a similar result.

Builders should now prioritize:

  • Down-scaling: Use smaller models (Terra/Luna) for high-volume, latency-sensitive, or repetitive tasks like data extraction.
  • Reasoning Tuning: Reduce "reasoning effort" settings. In production, GPT-5.6 at low effort often outperforms older models at high effort, compounding cost savings with improved latency.

Architectural Interventions for Efficient Agents

Beyond model selection, the Responses API now includes primitives that allow agents to operate with significantly fewer tokens while maintaining or increasing intelligence.

  • Programmatic Tool Calling: Instead of forcing the model to reason over large intermediate datasets, use programmatic tool calling to execute JavaScript. This allows the model to delegate data processing and filtering outside the context window, reserving its reasoning capacity for high-level judgment.
  • Multi-Agent Orchestration: Complex tasks should be decomposed. A primary agent can orchestrate subagents that work in parallel, passing results back for final synthesis. This approach is natively supported in the API and is highly steerable; developers can guide the model on when to spawn subagents to avoid unnecessary token expenditure.
  • Compaction and Retained Reasoning: These architectural interventions can dramatically improve performance on complex benchmarks. For example, using these features on ARC-AGI-3 resulted in a performance jump from 13.3% to 38.3% while simultaneously reducing output token usage by 6x.

Optimizing Context and Caching

To further improve efficiency, the prompt cache TTL has been extended to 30 minutes, and developers can now set deterministic cache breakpoints. By using an appropriate prompt_cache_key, developers can increase cache hit rates and ensure requests land on the same inference engine, which reduces latency and stabilizes performance across agentic workflows.