GPT-5.6 Builder's Guide: How OpenAI Rebuilt the Economics of AI Agents
On August 13, OpenAI released not a new model card but a construction manual for running agents cheaply. The GPT-5.6 family's pitch is a single phrase: price-performance.
The Cost Ledger
The most striking comparison: on BrowseComp, a fact-finding benchmark, GPT-5.5 (Extra High) scored 84.36% three months ago at a cost of $33.27. At release, GPT-5.6 Luna (Extra High) reached 84.04% for just $1.33. Hypha's engineering lead put it more bluntly: Luna preserved 98% of GPT-5.5's extraction accuracy at roughly 1/18 the cost.
More important is architecture-level savings. GPT-5.6 was trained end-to-end with a three-piece toolkit:
- Cross-turn reasoning persistence (carry reasoning across turns)
- Native compression (compact long conversations)
- Programmatic tool calling (move filtering, aggregation, and orchestration out of model context)
Where the Savings Stop
Cheapness has prerequisites. OpenAI repeatedly emphasizes turning down reasoning effort: GPT-5.6 Sol at the *low* tier already surpasses GPT-5.5 at the *high* tier, harness unchanged. Smaller models like Luna and Terra fit high-throughput, low-latency, repetitive steps inside agent pipelines. A legal-tech example: use them for handwritten memo extraction, then send only the structured output to a frontier model for analysis. Extended prompt-cache TTLs (at least 30 minutes) and deterministic cut points helped several startups cut un-cached input by 28%.
Cheap Is Not the Destination, It's the Inflection Point
What the guide is really saying is that agent economics have changed. Tasks that once required a frontier model at every step can now be handled with small models plus tuned reasoning effort plus architectural trade-offs, achieving equal or better results. When "expensive" is no longer the hard constraint on building agents, the bet shifts to who can design a more frugal harness. Cost has become the core competitive edge of the new generation of AI coding.