|
|
Welcome back! July is a good month to ship something, and we’d love to see what you build with everything new. OpenAI Build Week submissions are open until July 21 at 5 PM PT, and this issue is packed with GPT‑5.6 updates, new Work and Codex workflows, and practical ideas to help you get started.
|
|
|
|
|
What shipped, what changed, what to automate next
|
OpenAI Build Week registration is still open!
OpenAI Build Week (July 13- 21) is live and there’s still time to build and submit your project. Projects will be judged by OpenAI leadership: Thibault Sottiaux, Peter Steinberger, Kath Korevec, Tara Seshan, and Leah Belsky. Winners will be awarded $100k total in prizes, tickets to OpenAI DevDay, a free year of ChatGPT Pro, spotlights by OpenAI (like in this newsletter!) and more.
Across the week, join us at:
Build solo or bring a team. That idea sitting in your backlog now has a deadline.
|
|
|
GPT‑5.6: our new model family for complex production workflows
GPT-5.6 pairs stronger coding, tool and computer use, and design judgment with a new efficiency baseline. It’s now available in the API and Codex, with three models tuned for different production needs:
- GPT‑5.6 Sol: Our flagship model for the hardest coding and agentic work. At maximum reasoning, GPT-5.6 Sol outperforms Claude Fable 5 while using 54% fewer output tokens.
- GPT‑5.6 Terra: Our balanced model for everyday production workloads, with performance competitive with GPT‑5.5 at a lower price.
- GPT‑5.6 Luna: Our fastest and most affordable model for well-defined, high-volume work.
|
|
|
GPT-5.6 Sol sets a new industry standard for model capability, efficiency, and alignment, delivering more intelligence from every token. For the API, start with the GPT-5.6 model guide and migration quickstart, or ask Codex to run $openai-docs.
|
|
|
Meet the new ChatGPT: Chat, Work, and Codex
- Chat: the fastest path for getting answers, brainstorming, writing and everyday tasks.
- Work: for multi-step workflows, available on web, mobile, and desktop. Powered by the Codex harness and agent runtime, Work can operate across connected apps and files and turn a goal into finished materials.
- Codex: for repo-level development with local code, terminals, and developer tools on the desktop (and Codex on mobile to keep that work moving from anywhere).
In short: Chat for answers. Work for workflows. Codex for development.
|
|
|
|
|
|
|
Guides, blogs, and things worth cloning
|
Four tips to get the most of your token usage
- Cache the context you keep paying for: Reuse stable prompt prefixes instead of processing the same context on every request. Explicit prompt caching can cut repeat cost and latency.
- Preserve reasoning across turns: When goals, assumptions, and priorities stay stable, set reasoning.context to all_turns and continue with previous_response_id so GPT‑5.6 can use compatible reasoning from earlier responses.
- Let models compose and run JavaScript that orchestrates tool calls: Programmatic Tool Calling with GPT-5.6 turns open-ended tool use into predictable stages and runs the code in a hosted sandbox to keep it out of context.
- Audit your Codex instructions: Audit Skills and AGENTS.md for duplicated or overly prescriptive guidance. GPT‑5.6 follows intent more reliably, so slimmer instructions can preserve the workflow while reducing repeated context.
🎥 Learn more about Valuemaxxing with GPT-5.6 at our upcoming Build Hour, July 23, 2026, 10–11 AM PDT.
Do more all in one place with Codex
Keep long-running work moving across Codex threads
- Set the finish line with /goal: When the finish line is clear but the path isn’t, use a Goal in Codex to define the outcome, constraints, and proof of completion. Add a review, repair, and validation loop so Codex can test, fix, and continue until the checks pass.
- Split independent work across subagent threads: Use subagents in Codex to delegate exploration, tests, or implementation in parallel, steer them while they run, and return distilled results to the main thread. Try it for performance tuning, migrations, flaky-test investigations, or multi-step refactors.
Build a ChatGPT Workspace Agent, then trigger it from your app
- First, build a Workspace Agent that can use connected apps, follow a skill, run on a schedule, and save its output.
- Then, add an API trigger so a form, ticket, CRM record, or backend event can start the workflow. The run happens asynchronously with the result written to the agent’s configured destination.
|
|
|
|
|
Go behind the scenes with OpenAI engineers
|
Core dump epidemiology: fixing an 18-year-old bug
After analyzing a year of Rockset core dumps, OpenAI engineers found that two unrelated problems (a faulty Azure host and an 18-year-old race condition in GNU libunwin) were causing crashes that looked identical.
"Finding the libunwind race condition was a burst of dopamine, because the impossible puzzle pieces suddenly fit together, but my favorite part of this bug hunt was that it led me into new parts of the system. I got to learn about parts of C++ and Linux that were previously just black boxes to me."
—Nathan Bronson, Member of Technical Staff
Turn expert corrections into a self-improving agent
Expert corrections are most useful when they make the next version better; so this team built a way to turn production edits into targeted evals and scoped engineering tasks with explicit regression checks.
“Once the product could trace a practitioner’s correction, turn it into a targeted eval, and give Codex a bounded task, production signals stopped disappearing into a backlog. That’s when the system started to feel genuinely self-improving.”
—Arthur Fernandes Araujo & John de Wasseige, Members of Forward Deployed Staff
|
|
|
|
|
From side projects to professional work
|
Sol builds and trains a language model from one prompt
Pietro Schirano gave GPT-5.6 Sol one prompt and his iMessage archive. Sol built the local training pipeline and trained a 1.38M-parameter, four-layer decoder-only Transformer on 8M tokens, then generated replies in his writing style. Watch the build.
Sol helps get Vercel’s data science agent unstuck
Using GPT‑5.6 and Codex, Andrew Qu, Vercel’s Chief of Software, pushed an internal data science agent through work that had stalled earlier models. Hear how he did it.
A tower-defense game, one crisp Sol iteration at a time
Shuang Zheng used GPT-5.6 Sol and Codex to build Acornado, a playable tower-defense game. Short, crisp iterations produced heart effects, sound cues, and solid code coverage without over-engineering. Play Acornado and read the build notes.
Turning restaurant order data into a reusable Skill
At a Korea Codex Skillathon, first-place winner Jei-him Choi packaged a real restaurant table-ordering workflow as a Codex Skill. Feed it order CSVs; it returns store-level insights, mobile-friendly reports, shareable links, and email drafts for sales and customer success. Clone the winning Skill.
Make canine blood-donation programs easier to find
Inspired by her dog Graham (and a custom Codex Pet) Gabrielle Zhou built a prototype of an app called Good Dog Blood at a Women Who Codex event. The prototype explains canine blood donation and gives owners one place to find nearby donor programs. Watch the demo.
Giving Korea’s Codex community a home
Using Sites, Codex, and Annotations, Seowoo Han turned Korea’s community events into a public archive—then paired it with beginner-friendly resources for first-time Codex and GitHub users. Browse the archive, watch the walkthrough, or read the non-developer GitHub guide.
Built something worth sharing? Submit it to the OpenAI developer showcase.
|
Happy building,
The OpenAI Team
|
|
|
|
|
|
|
|
|
