OpenAI DevDay 2026 global tour: times, venues, and Codex/API

In one line: once the flagship day and the eight-city tour sit on the same calendar, the practical question is whether this week is for the livestream or for one city application. The sections below split that into local start times for five cities, then the Codex and Agents API surface that is already public—without turning an unpublished keynote into a feature list.

Two events, not one

OpenAI DevDay is Tuesday, September 29, 2026 at Fort Mason in San Francisco. It is for invited engineers, technical founders, researchers, and technical leaders. The site says in-person applications are closed. The opening keynote is livestreamed for free; other sessions are recorded and posted on openai.com afterward.

DevDay Exchange follows. The official events page lists Bengaluru on October 16, Tokyo on October 20, Seoul on October 22, Berlin on October 26, Paris on October 28, London on November 3, São Paulo on November 6, and Mexico City on November 10. Applications are still being reviewed and can still be submitted. Each event starts at 2 p.m. local time. Venue and full program are sent only after approval. A community post dated Mexico City as November 11; this article follows the events page, which says November 10.

San Francisco, September 29

The site lists the day in Pacific Time (PDT, UTC−7). U.S. daylight saving is still in effect on September 29. Invited registration is $650. Tickets are non-transferable, and there is no door sale.

Lunch and breakouts overlap on the official clock. That overlap is on the site, not a conversion error. If you are remote, the session worth scheduling is the 10:00 a.m. PDT keynote. Breakouts come later as recordings.

Tokyo, Seoul, Berlin, Paris, London

Exchange has one published rule: start at 2 p.m. local time. The European stops fall after the 2026 clock change. The EU and the UK both set clocks back on Sunday, October 25, so Berlin on October 26 and Paris on October 28 are CET (UTC+1). London on November 3 is GMT (UTC+0). Tokyo and Seoul stay at UTC+9.

All five are weekdays: Tokyo Tuesday, Seoul Thursday, Berlin Monday, Paris Wednesday, London Tuesday. OpenAI describes Exchange as technical sessions with product and engineering, real projects and Codex workflows, hands-on guidance on your own work, and open conversation with the OpenAI team. It is not a scaled-down copy of the San Francisco conference.

Codex and API already shipping

The Exchange page names Codex workflows as something attendees will see. That is not the same claim as “the keynote will ship a specific unannounced feature.” The list below is limited to what OpenAI had already published by September 24. If September 29 adds a model, price, or endpoint, use that day’s notes, not this page, as the changelog.

Codex as an embeddable harness

The August 19 developer post describes Codex as an open agent harness: conversation state, streamed execution, tools, sandbox and approval policy, and work that continues across turns. The CLI, app-server, and official SDK are open source. Three integration shapes:

• codex exec — scripts, CI, or a one-off background job that returns structured output
• Codex SDK — application code that starts, resumes, or streams a task
• app-server — a product that keeps a thread, receives events, interrupts work, exposes tools, and handles approvals

Read that next to Codex inside ChatGPT Work. Exchange sessions are more likely to be about attaching the workflow to a real repository than about desktop buttons.

Agents API: the same harness, hosted

On September 10 the Agents API entered public beta. OpenAI’s description: the harness and infrastructure behind Codex, exposed as an API. During the beta there is no extra fee for the API itself; you pay for tokens and tools. A session is created with a task, model, tools, and environment. The published example uses gpt-6-astra and shows MCP, subagents, and a vault.

• Context compaction — earlier context is compacted as a session nears its limit, so work can span windows
• Tool search — tool definitions load when needed, which cuts tokens and protects cache
• Programmatic tool calling — parallel calls, chained operations, and filtering in code before results return to context
• Subagents — the main agent splits work; each subagent keeps its own context

The environment can be an OpenAI-hosted sandbox, your own infrastructure, or a partner sandbox. Named partners include Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. The hosted sandbox is the same isolation layer used by Codex and ChatGPT, and it can be configured with files, packages, skills, and plugins.

What to do before the keynote

The keynote is 17:00 UTC on September 29. Three notes made the same day are more useful than staying up for every breakout.

1. Write down the current path. CLI, SDK, app-server, or Agents API—and which of those may touch a production repo.
2. Bring one small repository failure. Exchange is pitched as hands-on. Without a concrete miss, the session stays a product tour.
3. Keep macOS verification off the agent host. If the task includes an iOS build, confirm runner labels, certificates, and branch protection before an agent pushes.

# 1. Entry: CLI / SDK / app-server / Agents API
# 2. Default model, and whether gpt-6-astra is already in use
# 3. Tools: MCP, built-in search, or custom functions
# 4. macOS steps still on a separate runner
runs-on: [self-hosted, macOS, cloud-mac]

FAQ

What time do Tokyo, Seoul, Berlin, Paris, and London start? 2 p.m. local: 05:00 UTC for Tokyo and Seoul, 13:00 UTC for Berlin and Paris, 14:00 UTC for London. Dates are in the table above.

Where are they? Only Fort Mason is public. The five Exchange venues are sent after approval.

What if the San Francisco application is closed? Watch the free keynote on September 29. Breakout recordings follow. Exchange cities need their own application.

Do Codex features wait for DevDay? The open harness (August 19) and the Agents API beta (September 10) are already available. DevDay may show how teams use them, and it may ship something new. Endpoints that are not in an OpenAI post are not listed here as shipped.

Does the Agents API add a fee? The beta post says there is no extra API fee; you pay for tokens and tools. The San Francisco ticket is separate: $650 if you are invited.

• Using Codex after it moved into ChatGPT Work
• Another harness: deploying DeepSeek Harness
• Where a long-running coding agent should execute