GPT-5.6 Sol · workflow
Separate judgment from deterministic processing, then combine programmatic tool calls, parallel subagents, and prompt-cache boundaries into a long-running workflow whose cost, latency, citations, and failures can be reviewed.
Produce an auditable long-running workflow: the model handles judgment, while deterministic work stays in programs; every subagent, tool call, and cache boundary exposes inputs, outputs, latency, and failure state. This is an editorial workflow based on OpenAI’s builder guide, not a claim that Tabbit executes it with one click.
Confirm the Responses API and list the allowed tools with their schemas.
Separate the task into independently checkable judgment and deterministic steps.
Define human approval points, retry limits, a budget ceiling, and final acceptance criteria.
Prepare repeatable representative tasks; one successful run cannot establish a production success rate.
Start with the smallest model and prompt that can route the task; record why Sol is needed and the expected quality, cost, and latency.
Keep database queries, file handling, and format conversion in programmatic tool calls; do not make the model repeatedly copy large tool results.
Parallelize only independent work with clear deliverables; give each subagent an input boundary, output schema, and timeout.
Preserve sources, tool status, failure reasons, and human checkpoints during synthesis.
Put stable instructions, tool schemas, and invariant context in the cache prefix; record cache hits, cost, and latency changes.
Every stage has structured output traceable to its inputs, tool results, and sources.
Failed calls are not silently treated as successes; the final result lists missing evidence.
Parallel tasks cannot overwrite one another, and retries do not create duplicate side effects.
Record model, reasoning effort, tools, sample count, cost, and latency; label unmeasured values as unknown.
Human approvals and stop conditions are visible, and the final artifact meets the predeclared criteria.
Preserve an empty or malformed tool result, retry once, then hand off or use an explicit fallback.
If a subagent violates its schema, request a corrected output or mark it failed instead of asking the synthesizer to guess.
If cache hits fall, check whether dynamic content broke the prefix; do not hide necessary version information for a higher hit rate.
If cost or latency exceeds the limit, reduce context, parallel scope, or effort on non-critical steps.
Production examples and cost charts in the official guide are not your harness measurements; do not promise price, success rate, or latency.
Parallel agents add coordination and failure surfaces; do not split work merely to create more agents.
A tool call indicates a request or prepared action, not permission; high-risk operations still need human approval.
This official guide for builders of production applications covers model selection, the Responses API, Programmatic Tool Calling, multi-agent orchestration, and Prompt Caching, and offers recommendations on Sol’s cost efficiency and use for long-running tasks.
The following is the visible page text extracted during this visit. It includes page navigation, automatic translations, advertisements, comments, and other page elements; verify against the original URL before citing.
Skip to main content Research Products Business Developers Company Foundation (opens in a new window) Log in Try ChatGPT (opens in a new window) The builder’s guide to GPT‑5.6 | OpenAI
August 13, 2026
Applied AI The builder’s guide to GPT‑5.6
Technical lessons from startups in production
Listen to article 6:26 Share GPT-5.6 sets a new standard for price-performance A better out-of-the-box experience Model Selection Evolving the Responses API to architect more efficient agents Programmatic Tool Calling Multi-agent Prompt Caching Conclusion GPT‑5.6 sets a new standard for price-performance
The GPT‑5.6 model family makes frontier-level agent performance dramatically more affordable, while also advancing the frontier of what is possible.
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
Since GPT‑5, each model generation has sought to tackle longer-horizon tasks with fewer tokens. GPT‑5.6 continues that trajectory: stronger agent performance, lower costs, with minimal changes to the underlying harness.
Artificial Analysis Coding Index Cost Latency Output tokens Cost $0 $1,000 $2,000 $3,000 40 50 60 70 80 Index score API cost (USD) GPT-5.6 Sol GPT-5.6 Terra GPT-5.6 Luna GPT-5.5 Claude Fable 5 Claude Opus 4.8 Gemini 3.1 Pro Preview
The improvements in top-line cost efficiency are compounded with increased accuracy at lower reasoning efforts. For example, on Agents’ Last Exam, GPT‑5.6 Sol at “low” reasoning outperformed GPT‑5.5 at “high” reasoning when the harness was kept constant. We’ve seen similar success stories in production testing where startups report seeing significant cost improvements across a range of workflows by reducing the reasoning effort from the prior defaults.
“ We dropped GPT‑5.6 into our harness, and low reasoning effort gave us our best results. It knew when the data just wasn’t there, didn’t chase bad leads, and got to the right answer with fewer tokens.” — Izzy Miller, AI Research Lead, Hex (opens in a new window) 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 been in large part due to these models being significantly more capable than cost-optimized models at handling longer contexts and tool calling. 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.
1 of 3 “ 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 (opens in a new window) Hypha Browser Use PlayerZero
Consider tasks in BrowseComp: a search-based benchmark that tests a model’s ability to search for obscure facts. Three months ago, GPT‑5.5 (Extra High) scored 84.36% on this benchmark for a total cost of $33.27. At launch, GPT‑5.6 Luna (Extra High) delivers essentially the same performance, scoring 84.04% at a cost of $1.33. We’ve since reduced prices further. Read more on our latest price cuts.
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, if you’re operating a legal-tech startup that parses handwritten memos prior to agentic analysis, instead of using a frontier model for the entire use case, you can now use Terra or Luna for extraction and register significant cost savings.
Evolving the Responses API to architect more efficient agents
In addition to making GPT‑5.6 more performant out of the box, we also shipped new primitives to the Responses API to unlock further gains. We trained GPT‑5.6 end-to-end with three complementary architectural interventions that enable agents to operate more efficiently:
Reuse work already performed: by allowing reasoning to be persisted (opens in a new window) across model turns and using native compaction (opens in a new window) to compress long-running conversations, the model can maintain coherence in its work across longer task horizons without getting confused or having to reconstruct prior context. Parallel decomposition where appropriate: using native multi-agent orchestration (opens in a new window) allows coordinating multiple agents across parallel workstreams to finish complex tasks faster. Move deterministic work into code: using programmatic tool calling (opens in a new window) to filter, aggregate, and orchestrate tool outputs outside the model’s context window, reserving model tokens for judgment and reducing cost, latency, and context rot.
Used together, the difference can be dramatic. For example, on ARC-AGI-3, GPT‑5.6 Sol scored 13.3% with the standard harness. After enabling retained reasoning and compaction, however, the score jumped to 38.3%—while using roughly 6× fewer output tokens. No changes to the model, but nearly three times the performance. You can read more on our ARC-AGI-3 harness investigation here.
Programmatic Tool Calling
Agentic workflows often involve two kinds of work:
Tasks that require judgment 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 their outputs outside the context window. The model is left to focus on what requires intelligence: applying judgment.
“ For financial research, the hard part is reliably pulling filings, coordinating tools, and working through the numbers. In our evaluations, GPT‑5.6 using Programmatic Tool Calling matched our rubric quality while using 21% fewer input tokens. That’s the difference between an agent that can discuss financial research and one that can actually carry it out.” — Alex Wang, Applied AI, Rogo (opens in a new window) Multi-agent
On complex, parallelizable tasks, distributing actions and reasoning across multiple agent workstreams enables faster task completion as well as higher intelligence. In these setups, the primary agent is responsible for orchestrating the subagents and delegating tasks to them. The subagents pursue their objectives in parallel and finally pass back their output to the primary agent for final synthesis. Teams can start leveraging multi-agent natively by enabling multi-agent (opens in a new window) in the Responses API. This is also how the ultra capability setting in ChatGPT works.
“ Qualia runs teams of agents on open-ended research problems, and GPT‑5.6 Sol just clicked. It showed a marked improvement over GPT‑5.5, finished faster than almost every other model we tested, and quickly became our go-to OpenAI model.” — E Chi, Founder, Quadrillion (opens in a new window) “ GPT‑5.6 is the best orchestrator we’ve seen from OpenAI. We threw six specs at it at once (writing, building, and talking through all of them) and it kept track of everything without the quality falling apart.” — Jon Bell, Co-founder and CPO, Obvious (opens in a new window)
Although GPT‑5.6 has a strong sense of the appropriate number of subagents and when to spawn them, multi-agent behavior is very steerable. Instructing the model on when to invoke subagents can increase the likelihood of spawning agents only in situations where the additional token expenditures would result in better performance.
Prompt Caching
Across the entire family of models, 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. This has enabled startups to significantly improve their cache hit rate.
“ We added cache breakpoints and workspace-specific keys to a shared 29,000-token prompt and cut uncached input by 28%. The 30-minute cache window was a big unlock too: our agents could reuse the same context across runs instead of starting from scratch.” — Lorenzo Gentile, AI Engineer, Ploy (opens in a new window)
In addition to setting cache breakpoints, continuing to use an appropriate prompt_cache_key (opens in a new window) increases the likelihood of requests landing on the same inference engine as one that previously served the same prefix, thereby reducing latency.
Conclusion
What stands out across these examples is how much the economics of building agents have changed.
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.
We're excited to see what you all build!
2026 API Platform About the authors
This guide was developed by Samarth Madduru (opens in a new window) , Prashant Mital (opens in a new window) , Dave Leo (opens in a new window) , and Julien Reiman (opens in a new window) , based on their experience working closely with startups building on GPT‑5.6 from early testing through production.
Keep reading View all How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery
Applied AI Jun 23, 2026
Using AI to help physicians diagnose rare genetic diseases affecting children
Applied AI Jun 18, 2026
How an astrophysicist uses Codex to help simulate black holes
Applied AI Jun 11, 2026
Research Research Index Research Overview Economic Research Latest Advancements GPT-5.6 GPT-5.5 GPT-5.4 Safety Safety Approach Deployment Safety (opens in a new window) Security & Privacy Trust & Transparency Products ChatGPT (opens in a new window) ChatGPT Business (opens in a new window) ChatGPT Enterprise (opens in a new window) ChatGPT for Education (opens in a new window) Codex Release Notes API Platform Overview API Log In (opens in a new window) Docs (opens in a new window) Business Overview Solutions Resources Customer Stories Partner Network Contact Sales Developers Apps SDK (opens in a new window) Open Models Docs (opens in a new window) Resources (opens in a new window) Developer Forum (opens in a new window) Company About Us Our Charter Careers News Support Help Center (opens in a new window) More Stories Academy Supply Co. Livestreams Podcast RSS Terms & Policies Terms of Use Privacy Policy Other Policies (opens in a new window) (opens in a new window) (opens in a new window) (opens in a new window) (opens in a new window) (opens in a new window) (opens in a new window) OpenAI © 2015–2026 Your privacy choices English United States
OpenAI · Source date: 2026-08-13 · Edited: 2026-09-20
Read the original sourceGPT-5.6 Sol
Run this guide in the environment listed above. Downloading does not transfer the template or establish model availability for your account.
