GPT-5.6 Slashes Agent Costs 96%: Same Benchmark for $1.33 vs $33.27

OpenAI’s GPT-5.6 family is doing something genuinely unusual in frontier AI: getting cheaper and more capable at the same time. On BrowseComp, the search-heavy benchmark that tests a model’s ability to track down obscure facts, GPT-5.5 running at Extra High reasoning scored 84.36% at a cost of $33.27. Three months later, GPT-5.6 Luna hits 84.04% at the same reasoning effort for $1.33. That’s a 96% cost reduction for essentially identical performance. And OpenAI says prices have dropped further since launch.

The pattern holds across the family. On Agents’ Last Exam, GPT-5.6 Sol at “low” reasoning outperformed GPT-5.5 at “high” reasoning when the harness was kept constant. That’s a fundamental shift in how developers should think about model selection. The old playbook—always reach for the flagship at maximum reasoning for long-horizon tasks—no longer holds. Luna and Terra, the smaller models in the family, can now handle longer contexts and tool calling at a level that previously required GPT-5.4 or 5.5. For high-volume workloads, latency-sensitive interactions, or repeated agentic steps, the economics have flipped.

The bigger story, though, is what OpenAI shipped alongside the model. Three new primitives in the Responses API change how agents actually operate: persisted reasoning across turns, native conversation compaction, and programmatic tool calling that moves deterministic work into JavaScript running outside the context window. The results are striking. On ARC-AGI-3, GPT-5.6 Sol scored 13.3% with a standard harness. After enabling retained reasoning and compaction—no model changes—it jumped to 38.3% while using roughly 6× fewer output tokens. Nearly triple the performance by changing how the model is invoked, not the model itself.

That last number deserves attention. Context rot—the degradation that happens when models process too much intermediate data—has been a quiet killer of long-running agent workflows. Letting models offload filtering, aggregation, and orchestration to code rather than reasoning over every intermediate result in-context isn’t just a cost play; it’s an architectural fix for a failure mode that’s been plaguing agent deployments since the beginning. Multi-agent orchestration, where a primary agent delegates to parallel subagents, pushes in the same direction. The takeaway for builders is clear: the next wave of agent efficiency gains may come less from bigger models and more from smarter harnesses around them.