OpenAI’s GPT-5.6 model family delivers frontier-level agent performance at dramatically lower costs, with architectural improvements that reduce token usage and enable more efficient multi-agent orchestration. The smaller models in the family—Luna and Terra—now match or exceed the performance of previous flagship models on key benchmarks while costing orders of magnitude less. For example, on BrowseComp, GPT-5.5 Extra High scored 84.36% at $33.27, while GPT-5.6 Luna Extra High scored 84.04% at just $1.33. On Agents’ Last Exam, GPT-5.6 Sol at “low” reasoning outperformed GPT-5.5 at “high” reasoning under the same harness.

Three new architectural primitives in the Responses API unlock further efficiency: retained reasoning and compaction allow the model to persist and compress reasoning across turns; native multi-agent orchestration enables parallel subagent workstreams with steerable delegation; and programmatic tool calling moves deterministic work (filtering, aggregation, orchestration) into JavaScript, reserving model tokens for judgment. Combined, these interventions yielded dramatic results: on ARC-AGI-3, GPT-5.6 Sol scored 13.3% with the standard harness but jumped to 38.3% after enabling retained reasoning and compaction—using roughly 6x fewer output tokens.

Prompt caching has also improved: the cache TTL is now a minimum of 30 minutes, and cache breakpoints can be set deterministically within the context window. The post recommends using smaller models like Luna or Terra for high-volume, latency-sensitive extraction tasks, and only invoking higher reasoning effort when judgment-intensive work is required. The broader message is that the economics of building agents have shifted: use cases that once required a frontier model at every step can now achieve comparable or better results at a fraction of the cost by combining smaller models, tuned reasoning effort, and efficient architectural choices.