Glonce

OpenAI's GPT-5.6 guide shows agents can match frontier results at a fraction of the cost

agentsThe Glonce editorsOpenAI🔥67Glonce rating67Impact60Novelty55Relevance100Surprise25Players95Credibility90

OpenAI has published a builder's guide to its GPT-5.6 model family, walking through how startups are combining smarter model selection with new Responses API features to build AI agents that cost far less to run. The guide frames GPT-5.6 as a continuation of a trend that started with GPT-5, where each generation handles longer-horizon agent tasks with fewer tokens; it says GPT-5.6 continues that trend with only minimal changes to the underlying agent harness.

The GPT-5.6 family includes a larger model called Sol, alongside two smaller, cheaper models called Luna and Terra. OpenAI says that with more test-time compute, Luna and Terra can often perform similarly to the previous-generation GPT-5.4 and GPT-5.5 models while costing significantly less. Historically, the guide says, only a flagship model at the highest reasoning setting was the reliable choice for long-horizon tasks; it says that has changed with the 5.6 family. The guide backs this with a specific case: on BrowseComp, a benchmark that tests a model's ability to search for obscure facts, GPT-5.5 at "Extra High" reasoning scored 84.36% three months earlier, at a total run cost of $33.27. At GPT-5.6's launch, Luna at "Extra High" reasoning scored 84.04%, essentially the same result, at a run cost of $1.33, roughly a 96% drop from the earlier $33.27 figure. OpenAI says Luna's price has been cut further since that figure, without giving a new number. The guide recommends the smaller 5.6-family models for high-volume workloads, latency-sensitive interactions and repeated steps inside an agent pipeline; as a hypothetical illustration, not a named case study, it describes a legal-tech startup that parses handwritten memos before agentic analysis using Terra or Luna for the extraction step instead of running a frontier model through the whole pipeline, for significant cost savings.

Alongside model selection, OpenAI says it trained GPT-5.6 end-to-end with three architectural changes to the Responses API meant to make agents more efficient. The first lets the model reuse work already done: reasoning can now be persisted across turns, and native compaction compresses long-running conversations, so the model keeps coherence across longer task horizons without getting confused or reconstructing prior context each time. The second is native multi-agent orchestration, where a primary agent coordinates subagents that work in parallel and hand their output back for final synthesis; OpenAI says this is also the mechanism behind the "ultra" capability setting in ChatGPT. GPT-5.6 has what OpenAI calls a strong sense of when and how many subagents to spawn on its own, but the behavior is described as very steerable through instructions. The third change is programmatic tool calling, which lets GPT-5.6 write JavaScript to orchestrate tools, run independent calls in parallel and process their outputs outside the model's context window, reserving the model's own reasoning for the parts of a task that need judgment. As a hypothetical example, the guide describes an agent that retrieves 100 filings and needs to filter them by date and identify relevant transactions without reasoning over every intermediate result.

The guide's clearest demonstration of the combined effect comes from ARC-AGI-3: GPT-5.6 Sol scored 13.3% using the standard agent harness, and the same model, with no changes to the model itself, scored 38.3% once retained reasoning and compaction were enabled, while using roughly six times fewer output tokens. OpenAI's own summary: "No changes to the model, but nearly three times the performance." It points to a separate writeup of that ARC-AGI-3 harness investigation for more detail. A second comparison, on a benchmark called Agents' Last Exam, has GPT-5.6 Sol at "low" reasoning effort outperforming GPT-5.5 at "high" reasoning effort with the harness held constant, and OpenAI says startups in production testing have reported similar cost savings simply by lowering reasoning effort from their previous defaults.

The guide also covers prompt caching. Across the whole GPT-5.6 family, the prompt cache time-to-live has been extended to a minimum of 30 minutes, and cache breakpoints can now be set deterministically within a model's context window; OpenAI says this has let startups significantly improve their cache hit rates, though it gives no specific figure for that improvement. It also recommends continuing to use an appropriate prompt_cache_key, which raises the odds a request lands on the same inference engine that served the same prefix before, cutting latency.

OpenAI's own conclusion is that the economics of building agents have shifted: work that once needed a frontier model at every step can now get comparable or better results for a fraction of the cost by using smaller models, tuning reasoning effort and making more efficient architectural choices.

Key facts

  • GPT-5.6's smaller Luna model, at "Extra High" reasoning, scored 84.04% on the BrowseComp benchmark, essentially matching GPT-5.5's 84.36% from three months earlier, while the run's cost dropped from $33.27 to $1.33.
  • Enabling GPT-5.6 Sol's new retained-reasoning-and-compaction settings, with no change to the model itself, raised its ARC-AGI-3 score from 13.3% to 38.3% while using roughly six times fewer output tokens.
  • OpenAI trained GPT-5.6 end-to-end with three new Responses API primitives: persisted reasoning across turns with native compaction, native multi-agent orchestration, and programmatic tool calling that lets the model write JavaScript to run tools outside its context window.
  • Across the GPT-5.6 family, prompt cache time-to-live has been extended to a minimum of 30 minutes, with cache breakpoints now settable deterministically within the context window.
  • On Agents' Last Exam, GPT-5.6 Sol at "low" reasoning effort outperformed GPT-5.5 at "high" reasoning effort with the harness held constant.

Why it matters

GPT-5.6 itself had already shipped before this guide went up; the guide's real content is a demonstration of how much cheaper agents built on it can be without losing accuracy. OpenAI frames this as a continuation of a trend since GPT-5, where each generation handles longer-horizon agent tasks with fewer tokens, and says GPT-5.6 does it with only minimal changes to the underlying agent harness. Its clearest example: on the BrowseComp benchmark, Luna essentially matched GPT-5.5's accuracy while cutting the run's cost by about 96%, from $33.27 to $1.33. Its most striking one: turning on two new settings, retained reasoning and compaction, took GPT-5.6 Sol's ARC-AGI-3 score from 13.3% to 38.3%, nearly tripling it with no change to the model at all. Those are concrete numbers for rethinking model and reasoning-effort choices, not just a generic claim about being cheaper.

Who it affects

Teams building agent products on OpenAI's API: anyone choosing between GPT-5.6's Sol, Luna and Terra models for a given workflow, or using the Responses API's persisted reasoning, multi-agent orchestration or programmatic tool calling features directly. The guide's own illustration, a legal-tech startup parsing handwritten memos before agentic analysis, is explicitly hypothetical, not a named customer, but stands in for any startup running high-volume or repeated extraction-style steps inside a longer agent pipeline. It also touches ChatGPT users indirectly: OpenAI says the same native multi-agent orchestration under the hood powers ChatGPT's "ultra" capability setting.

How to use it

OpenAI's own guidance: reserve the larger Sol model and higher reasoning efforts for steps that actually need them, and route high-volume, latency-sensitive or repeated steps, like data extraction, to the cheaper Luna or Terra models. Turn on persisted reasoning and native compaction so a long-running agent keeps context across turns instead of reconstructing it each time. Enable multi-agent orchestration in the Responses API for parallelizable tasks, and give the model explicit instructions on when to spawn subagents rather than relying only on its own judgment. Use programmatic tool calling, JavaScript specifically, to move filtering, aggregation and orchestration logic outside the model's context window so the model only reasons over the parts of a task that need judgment. For caching, keep using a consistent prompt_cache_key: the cache TTL is now at least 30 minutes across the whole family, and cache breakpoints can be set deterministically within the context window, both of which push up the cache hit rate and cut latency.

How solid is it

This is a first-party OpenAI post, not an independent benchmark review, so every figure comes from OpenAI's own runs on its own harness. Where it gives numbers, they are precise: exact accuracy percentages and exact dollar run costs for BrowseComp, and for ARC-AGI-3, exact accuracy percentages plus an approximate token-reduction figure, roughly six times fewer output tokens, which the source itself hedges as "roughly." But the broader Model Selection claim, that Luna and Terra "can often perform similar" to GPT-5.4 and GPT-5.5, rests on just that one BrowseComp comparison, and that comparison is against GPT-5.5 only, not GPT-5.4. The "further" price cuts mentioned after the $1.33 figure come with no new number attached, and the claimed improvement in cache hit rate from the longer TTL and deterministic breakpoints is not quantified, only characterized as significant. The fetched guide also includes a chart or image carousel marked "1 of 3" in its Model Selection section; only that first panel's label came through, so any data in the other two panels is not reflected here.

Risks and caveats

Every efficiency figure here is OpenAI's own self-reported benchmark result on its own agent harness, not an independently verified one. The legal-tech extraction example is explicitly hypothetical, a stand-in for a workflow type, not a deployed case study, so real savings on a similar pipeline could differ. The guide gives no dollar cost for either ARC-AGI-3 run, so while the accuracy and token-count gains are quantified, the cost side of that specific example is not. Multi-agent orchestration's payoff also depends on how it is steered: OpenAI describes GPT-5.6's default judgment on spawning subagents as strong but the behavior itself as "very steerable," meaning outcomes depend partly on a team's own instructions, not purely on the model's defaults.

“No changes to the model, but nearly three times the performance.”

— OpenAI's GPT-5.6 builder's guide