OpenAI's official builder's guide does not offer a single “universal prompt.” Instead, it recommends designing workflows around model routing, reasoning-effort levels, preserving reasoning, native compaction, multi-agent systems, programmatic tool calling, and Prompt Caching. For Luna, the core recommendation is to use it for high-frequency, latency-sensitive, repetitive, and extraction tasks; more capable models can handle complex judgments.
Choose a model based on task risk and acceptance criteria; do not default to a flagship model at every step.
Make static system instructions, tool definitions, and reference materials form a stable prefix, then place the changing user question afterward.
Use multi-agent systems for independent workflows that can run in parallel; give judgment to the model, and leave filtering, aggregation, and data movement to code.
Use explicit completion checks instead of vague instructions such as “think forever” or “do it perfectly.”
Luna is well suited to high-frequency, latency-sensitive, repetitive, and document-extraction tasks. For final decisions or high-risk synthesis, define an escalation or human-review boundary.
Goal: [Clearly state the final result or system state to produce]
Context: [Provide sources, files, data, and prior decisions that could change the answer]
Output: [Specify the audience, format, depth, and evidence requirements]
Boundaries: [Specify facts that cannot be changed, actions that require approval, and stop conditions]
Completion check: [Specify what must be verified before completion]The following is the main visible body text extracted through the Tabbit international app; navigation and the footer have been omitted.
The builder’s guide to GPT‑5.6
August 13, 2026
Technical lessons from startups in production
GPT-5.6 sets a new standard for price-performance. The GPT‑5.6 model family makes frontier-level agent performance dramatically more affordable. In this guide, we show how startups are using smarter model selection and new API controls that help with reasoning continuity, multi-agent orchestration, and programmatic tool calling to build faster, more capable agents at a fraction of the cost.
A better out-of-the-box experience
GPT‑5.6 continues the trajectory of tackling longer-horizon tasks with fewer tokens: stronger agent performance, lower costs, with minimal changes to the underlying harness.
Model Selection
Historically, upgrading to a flagship model at the highest reasoning available has been the best option for long-horizon use cases. This has changed with the 5.6-family: with more test-time compute, Luna and Terra can often perform similar to GPT‑5.4 and 5.5 while being significantly cheaper.
“Luna keeps 98% of GPT‑5.5’s extraction accuracy at one-eighteenth the cost. That gives our agents high-quality document understanding at a price that makes it practical across many more workflows.” — Serhii Shchoholiev, Engineering Lead, Agents, Hypha
On BrowseComp, GPT‑5.5 (Extra High) scored 84.36% at a total cost of $33.27. At launch, GPT‑5.6 Luna (Extra High) delivered essentially the same performance, scoring 84.04% at a cost of $1.33. OpenAI says prices were reduced further afterward.
The smaller 5.6-family models are a strong fit for high-volume workloads, latency-sensitive interactions, and repeated steps within agentic workflows. For example, a legal-tech startup can use Terra or Luna for extraction before using a frontier model for agentic analysis.
Evolving the Responses API
OpenAI identifies three complementary interventions:
Reuse work already performed: allow reasoning to be persisted across model turns and use native compaction to compress long-running conversations.
Parallel decomposition where appropriate: use native multi-agent orchestration to coordinate multiple agents across parallel workstreams.
Move deterministic work into code: use programmatic tool calling to filter, aggregate, and orchestrate tool outputs outside the model’s context window, reserving model tokens for judgment.
Programmatic Tool Calling
Agentic workflows often involve tasks that require judgment and work that mostly requires moving, filtering, and combining data. When an agent retrieves 100 filings, filters them by date, and identifies relevant transactions, the model shouldn’t have to reason over every intermediate result in its context window. Programmatic Tool Calling lets GPT‑5.6 write JavaScript to orchestrate tools, run independent calls in parallel, and process outputs outside the context window.
Multi-agent
On complex, parallelizable tasks, distributing actions and reasoning across multiple agent workstreams enables faster task completion as well as higher intelligence. The primary agent orchestrates the subagents, delegates tasks, and finally synthesizes their output. OpenAI notes that multi-agent behavior is steerable: instructing the model when to invoke subagents can increase the likelihood of spawning agents only when the extra token expenditure is worthwhile.
Prompt Caching
Across the family, the prompt cache TTL has been extended to a minimum of 30 minutes and cache breakpoints can now be set deterministically within a model’s context window. Setting an appropriate prompt_cache_key increases the likelihood of requests landing on the same inference engine as one that previously served the same prefix, reducing latency.
Conclusion
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 using smaller models, tuning reasoning effort, and making efficient architectural choices.
GPT-5.6 Luna