OpenAI DevDay 2026 Prep — What Shipped This Week, What Got Documented, and What Got Leaked

OpenAI DevDay 2026 starts at 10:00 AM PDT on Sep 29 at Fort Mason. Here is the verifiable state of OpenAI's stack 24 hours before the keynote: GPT-6 Sol and Luna shipped Sep 22, the Agents API already documents sessions, multi-agent orchestration, and hosted sandboxes, and TestingCatalog has reported code and UI references for an always-on assistant called o.

September 28, 2026 – OpenAI’s fourth annual DevDay keynote starts at 10:00 AM PDT on Tuesday, September 29, at Fort Mason in San Francisco. The keynote runs roughly 90 minutes with Sam Altman on stage; the rest of the day is breakout sessions and a closing reception. What OpenAI actually announces tomorrow is not publicly known. What is publicly known, as of this morning, is the state of OpenAI’s stack across the last seven days. That state is the subject of this post.

Three things happened between Sep 22 and Sep 27 that any DevDay recap will be measured against: a same-day release collision between GPT-6 Sol and Anthropic Opus 5.5, three Agents API capabilities that quietly went live in OpenAI’s developer documentation, and a Sep 26 TestingCatalog report citing code and UI references for an assistant named “o.” Frame each one by what it actually is – shipped, documented, or reported – and the keynote ceiling becomes a lot easier to reason about.

What shipped: GPT-6 Sol and Luna, same day as Opus 5.5

On September 22, OpenAI released GPT-6 Sol and GPT-6 Luna in ChatGPT, the API, and Codex (MacRumors, Crowdfund Insider). Sol is the cheaper reasoning-tier slot; Luna is the cheaper non-reasoning tier. Both slot below the GPT-6 flagship on price and above the legacy GPT-5.6 line on capability.

That same day, Anthropic shipped Opus 5.5 (note.com weekly recap, Sep 27). Two frontier-grade model releases from the two largest US labs on the same calendar day is no longer a coincidence. Frontier-model cadence is now weekly, not monthly. If DevDay announces an “o” assistant or a new agent platform, the timing of Sep 22 says OpenAI already had the September release locked in – DevDay is the marketing surface, not the ship date.

What got documented: three Agents API capabilities already live

Three Agents API pages are live in OpenAI’s developer docs as of Sep 27, and they describe platform capabilities, not teasers:

None of these pages are dated as new, but the aiidelist.com tracker (last reviewed 05:50 AM PDT Sep 27) classifies all three as already-shipped platform capability, not announcement candidates. The Agents API has been in some form since April, but durable sessions + subagent delegation + hosted sandboxes together is the smallest set of features that lets a third party run a long-lived agent that can spin up subagents and execute code, all under OpenAI’s sandbox. That is the platform story tomorrow’s keynote is most likely to frame, regardless of what new product names come out of it.

What got leaked: TestingCatalog reports an “o” assistant

On Sep 26, TestingCatalog published code and UI references for an assistant named “o” inside the ChatGPT upgrade flow. The concrete evidence: an assistant-name string in the upgrade UI, a display-name config entry, and a “-o” email-suffix configuration. The headline frames it as “always-on agent”; the evidence base is code and UI strings, not launch confirmation.

The aiidelist.com tracker – the cleanest independent pre-DevDay evidence index I’ve found – explicitly tags this as “Reported leak,” not confirmed. OpenAI has made no announcement. The testingcatalog report does not establish a launch date, a working email feature, or an API surface. The Sep 22 GPT-6 Sol/Luna release and the three Agents API documentation pages are the documented state. The “o” assistant is a code-leak report, not a shipped product. Treat it accordingly until OpenAI ships or denies it from the Fort Mason stage.

One more pre-keynote signal: Tibo’s “ambitious sprint”

On Sep 24, Thibault Sottiaux (Tibo), an OpenAI team lead, posted on X about an “ambitious sprint” and “changes to how people work” (X mirror via twiscan). The post names no product. It is a teaser in the strict sense – a public signal from someone at the company that something is shipping this week. Pair it with the TestingCatalog report and you get a small triangulation: OpenAI staff are publicly excited, third parties are reporting code references, and the documentation surface is already live. That is the case for “something ships tomorrow,” not confirmation of what it is.

Evidence taxonomy: what is actually known

The aiidelist.com tracker separates four evidence statuses. Here is how the publicly known items land one day before the keynote:

Status Item Date Source
Confirmed (shipped) GPT-6 Sol and Luna in ChatGPT, API, Codex Sep 22 MacRumors
Confirmed (shipped) Anthropic Opus 5.5 Sep 22 note.com weekly recap
Confirmed (documented) Agents API sessions Live by Sep 27 OpenAI docs
Confirmed (documented) Agents API multi-agent (subagents) Live by Sep 27 OpenAI docs
Confirmed (documented) Agents API hosted environments Live by Sep 27 OpenAI docs
Confirmed (event) DevDay 2026 at Fort Mason, Sep 29, 10:00 AM PDT, livestreamed – devday.openai.com
Confirmed (event) DevDay Exchanges in 8 cities (Bengaluru, Tokyo, Seoul, Paris, Berlin, London, São Paulo, Mexico City) – devday.openai.com
Reported leak “o” always-on assistant – assistant-name string, display-name config, “-o” email-suffix config in ChatGPT upgrade UI Sep 26 TestingCatalog
Teaser Tibo post on “ambitious sprint” and “changes to how people work” Sep 24 twiscan mirror
Speculation DevDay keynote product announcements – None

Two notes on the table. First, the “Confirmed (event)” rows are taken from OpenAI’s own event page (devday.openai.com), which is a marketing surface but is still a primary OpenAI source for the schedule, the venue, the livestream, and the eight-city Exchange tour. Second, the Speculation row is empty on purpose – the keynote content is unknown. Do not invent product announcements.

What we do not know

We do not know what Sam Altman announces tomorrow. The “o” leak points at an always-on assistant with email-style identity, but code references are not a product spec. Tibo’s post is a mood, not a feature list. OpenAI’s documentation surface already covers sessions, subagents, and hosted sandboxes – which suggests an “agent platform” framing, not a single-product reveal – but the keynote could ship one of three shapes: a new consumer-facing assistant, a developer-facing agent runtime, or both. Treat any of those as equally possible until 10:00 AM PDT on Sep 29.

For the practical takeaway: if you build on the OpenAI API, the documentation is already your source of truth. Sessions, subagents, and hosted sandboxes are live; start there. If you build consumer products on top of ChatGPT, the Sep 22 GPT-6 Sol/Luna release is the floor under any new assistant that gets announced tomorrow. And if you read a “confirmed” headline about a DevDay product before the keynote ends, ask for the primary source. TestingCatalog reports code; OpenAI ships announcements. Those are different evidence classes.

The keynote is in roughly 22 hours. This is the floor. Tomorrow’s coverage starts there.

Sources