The GPT‑5.6 model family materially lowers the operating cost of agentic systems while advancing capabilities: smaller models can now handle longer‑horizon tasks with fewer tokens, and new Responses API primitives improve agent efficiency and orchestration.
Models and cost efficiency
Since GPT‑5, each generation has aimed to solve longer tasks with fewer tokens. GPT‑5.6 continues that trend: agent performance improves and costs drop, with minimal changes required to existing harnesses.
These headline cost gains are compounded by increased accuracy at lower reasoning effort. For example, on the Agents’ Last Exam benchmark, GPT‑5.6 Sol at “low” reasoning outperformed GPT‑5.5 at “high” reasoning when the harness remained constant. Startups testing in production report similar wins: reducing the prior default reasoning effort often yields significant cost improvements across workflows.
Historically, long‑horizon use cases typically required flagship models run at the highest reasoning setting because those models handled long contexts and tool calls far better than cost‑optimized variants. The 5.6 family changes that: with adequate test‑time compute, smaller models like Luna and Terra can often match GPT‑5.4 and 5.5 performance while being substantially cheaper.
Benchmarks and price example
On BrowseComp — a search‑based benchmark that measures ability to find obscure facts — GPT‑5.5 (Extra High) scored 84.36% three months ago 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. Prices have been reduced further since.
When to prefer smaller models
The smaller models in the 5.6 family fit high‑volume workloads, latency‑sensitive interactions, and repeated steps within agentic pipelines. For instance, a legal‑tech startup that parses handwritten memos before agentic analysis can use Terra or Luna for extraction instead of a frontier model for the whole flow, registering notable cost savings.
New Responses API primitives and architectural interventions
Alongside out‑of‑the‑box improvements, new Responses API primitives were shipped to unlock additional gains. GPT‑5.6 was trained end‑to‑end with three complementary architectural interventions that enable more efficient agent operation:
- Retained reasoning: preserving relevant chains of thought across steps so reasoning need not be repeated from scratch.
- Compaction: compressing model outputs into denser representations that retain crucial information using fewer tokens.
- Programmatic Tool Calling: letting the model write JavaScript to orchestrate tools, run independent calls in parallel, and process outputs outside the context window.
Using these features together can be dramatic. On ARC‑AGI‑3, GPT‑5.6 Sol scored 13.3% with the standard harness; after enabling retained reasoning and compaction the score rose to 38.3%, while consuming roughly six times fewer output tokens. The model itself did not change, but overall performance nearly tripled.
Programmatic Tool Calling and parallel processing
Agentic workflows typically separate raw extraction/processing from deliberative judgement. When an agent downloads 100 filings, filters by date, and identifies relevant transactions, the model doesn't need to reason over every intermediate result inside its context window. Programmatic Tool Calling enables GPT‑5.6 to generate JavaScript to orchestrate tools, run parallel calls, and aggregate results outside the context window, leaving the model to focus on judgment where intelligence is required.
Multi‑agent orchestration
For complex, parallelizable tasks, distributing actions and reasoning across multiple agent workstreams yields faster completion and often higher overall intelligence. In these setups a primary agent orchestrates and delegates to subagents; subagents pursue objectives in parallel and return outputs for final synthesis. Teams can enable native multi‑agent functionality in the Responses API; this is the same approach behind the ultra capability setting in ChatGPT.
Although GPT‑5.6 has a strong internal sense of how many subagents to spawn and when, multi‑agent behavior is highly steerable: instructing the model about when to invoke subagents raises the likelihood that spawn events occur only when the extra token spend will improve performance.
Prompt caching and deterministic cache breakpoints
Across the family, 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. Startups have leveraged this to materially improve cache hit rates. In addition, continuing to use an appropriate prompt_cache_key increases the likelihood that requests land on the same inference engine that previously served the same prefix, reducing latency.
Conclusion
These examples illustrate how the economics of building agents have shifted. Workflows that once required a frontier model at every step can often achieve comparable or better results at a fraction of the cost by using smaller models, tuning reasoning effort, and applying efficient architectural choices.
This guide was developed by Samarth Madduru, Prashant Mital, Dave Leo, and Julien Reiman based on their experience working closely with startups building on GPT‑5.6 from early testing to production. The authors express enthusiasm for what the community will build with these tools.
