OpenAI DevDay 2026: ChatGPT, Codex and Tools to Watch
OpenAI is scheduled to open DevDay 2026, its annual developer conference, in San Francisco on September 29, with the keynote at 1PM ET / 10AM PT. CEO Sam Altman is expected to take the stage alongside other company executives. Engadget OpenAI
The company describes DevDay as a working event for developers, built around technical sessions, hands-on demos, workshops, and time with the teams building developer tools. OpenAI Those sessions cover API behavior, default settings, and integration paths for adding the technology to software.
The stated agenda for this year is a first look at what is next in ChatGPT, Codex, and tools for building. OpenAI Live coverage is organized around the same split, with Engadget planning its live blog around ChatGPT and Codex announcements from the keynote. Engadget
To put that agenda in practical terms, each item covers a different point of contact. ChatGPT is the assistant that end users talk to. Codex is the coding system that works in code repositories, terminals, and CI, the automated systems that test and build software. Tools for building are the parts underneath, the endpoints, or connection points, SDKs, or reusable code kits, and policy controls that let teams run the other two in production.
In my view, the questions that matter are unglamorous. Teams building on large models depend on versioning, migration cost, latency budgets, or response-time limits, context handling, or how much text the model can keep in mind, tool-use reliability, permission scopes, observability, or tracking what the system does, and eval harnesses, or repeatable tests. A new capability only changes plans if inputs and outputs stay stable, failure modes are clear, and cost under load is predictable. Keynotes tend to show the smooth path. Production decisions wait for documentation, rate limits, deprecation policy, and support in existing pipelines.
The broader context here is how developer platforms earn trust. The pattern has held from PCs to the web to mobile to cloud. Adoption follows when a vendor lowers the cost of trying things without raising the cost of maintenance. That requires clear boundaries between model, supporting code, and application code, debugging tools that show why an agent acted and not only what it made, and controls for data boundaries, review steps, and safe rollback. For Codex-style work, the usual checks are repository permissions, execution sandboxes, or isolated areas for running code, review queues, and audit trails.
Looking further out, there is reason for measured optimism about what better tooling allows. Shorter iteration cycles let small teams test integrations that once needed dedicated machine-learning infrastructure. More programmable assistant behavior lets product teams add domain logic without rebuilding the interaction for every edge case. That does not remove careful systems work. It shifts the work toward product definition, testing, and operations, where durable software value has usually been created.
