What OpenAI published on August 13 wasn't a new model card — it was a construction manual for running agents cheaply. The GPT-5.6 family's selling point boils down to four words: price-performance.
The Cost Math
The most striking comparison: on BrowseComp, a fact-searching benchmark, GPT-5.5 (Extra High) scored 84.36% at a cost of $33.27 three months ago. At launch, GPT-5.6 Luna (Extra High) achieved 84.04% for only $1.33. Hypha's engineering lead put it more bluntly: Luna retained 98% of GPT-5.5's extraction accuracy at one-eighteenth the cost.
Even more important is "architectural savings." GPT-5.6 was trained end-to-end with a three-piece toolkit:
- Persistent reasoning across turns (retaining reasoning)
- Native compaction (compacting long conversations)
- Programmatic tool calling (moving filtering, aggregation, and orchestration out of the model context)
The Limits of Cheap
Cheapness has prerequisites. OpenAI repeatedly stresses turning reasoning effort down: GPT-5.6 Sol at the *low* setting beats GPT-5.5 at *high* (same harness). The smaller Luna/Terra models suit high-throughput, low-latency, repetitive steps in agent pipelines — for example, legal-tech workflows that first extract from handwritten memos, then feed frontier-tier analysis. Prompt cache TTL extended to at least 30 minutes, with deterministic breakpoints, helped a batch of startups cut cache-miss input by 28%.
Cheap Isn't the End Point — It's an Inflection Point
What this guide is really saying: the economics of agents have changed. Tasks that once required a frontier model at every step can now match or beat results using small models + lower reasoning effort + architectural trade-offs. When "expensive" is no longer a hard constraint on building agents, the bet shifts to who can design the most cost-efficient harness — cost itself becomes a core competitive capability for the new generation of AI coding.