OpenAI's DevDay 2026 is on September 29 in San Francisco. As of September 25 the event page says "Save the date for September 29th in San Francisco", calls it OpenAI's biggest event of the year, and offers a form to be notified when applications open. That is all. There is no agenda, no product list and no keynote description. So this post is not a preview of announcements. It is a list of what OpenAI has already shipped on its API this month, what it switches off in the days around the event, and the work a team can finish before Tuesday so that whatever is announced lands on a tidy account.
The reason to do this now rather than after is simple. September brought an agent runtime in beta, a voice model at a per-minute price, two new controls over API keys, two new models and a model with unusual request restrictions. A team that has not absorbed those will spend the week after DevDay catching up on both.
- 01The event page states a date and a city. Everything else about DevDay is unconfirmed.We do not repeat rumoured model names or leaked prices. Our standard for that is the same one we set in our post on launch signals: name a signal as unverified or leave it out.
- 02Four legacy models shut down on September 28, the day before the event.gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 all point at gpt-5.6-terra on the deprecations page. The Sora 2 Videos API was removed on September 24.
- 03September's API changes are the real preparation list.Agents API public beta, GPT-Live 1 at $0.05 a minute, key-creation governance, key expiry limits, GPT-6 Sol and Luna, and GPT-6 Astra's restrictions on reasoning effort, sampling and tool calling.
- 04Six checks fit in the two working days left, and none of them depends on what is announced.Grep for retired model IDs, set a key policy, set an expiry maximum, confirm Responses API coverage, decide session handling for agents, and run one costed voice test.
01 — The factWhat the event page says, and what it does not
The DevDay 2026 page says, in full, "Save the date for September 29th in San Francisco", that "Our biggest event of the year is back", and "Sign up to be notified when applications are open". It does not list sessions, speakers, a keynote time or a livestream, and it says nothing about who is admitted. We re-read it on the morning of September 25 and will not characterise it beyond that. If the page changes before Tuesday, the change is the news, and it should be read on the page rather than here.
What the page does establish is a deadline. Whatever is announced will be described against the API as it stands on September 29, which is the API described in the next two sections.
02 — The baselineWhat OpenAI already shipped on the API in September
Every item below is from OpenAI's API changelog, read on September 25, with the changelog's own date. The most recent entry is dated September 22.
- GPT-6 Astra, and long-running work controlsThe changelog lists what Astra does not support: the none reasoning effort, custom temperature and top_p, logprobs, and tool calling through Chat Completions. The same day added async tool calling, mid-turn steering over WebSockets and mid-conversation effort changes on the Responses API.
- Sep 3
- Prompt cache diagnostics, generally availableCompare cache reuse against a previous response, for GPT-5.6 and later. Our post on this week's caching changes covers what moved after.
- Sep 8
- Agents API, public betaSession orchestration, context compaction and recovery, with durable sessions, streamed progress, your own tools and MCP servers, and agents that run in OpenAI-hosted sandboxes or a sandbox from your own infrastructure or a supported provider.
- Sep 10
- GPT-Live 1, generally availableFull-duplex voice at $0.05 per minute, with the backend model and tool usage charged separately. Two delegation modes: Responses or client.
- Sep 10Per-minute price
- API key expiration dates and maximum lifetimesProject keys can carry an expiry, and administrators can require every new key to expire within a limit set at the organisation or project.
- Sep 10
- API key creation governanceIn the changelog's words, administrators "can allow only service-account keys, allow only user-owned project keys, or disable all new API key creation". Existing keys are unaffected.
- Sep 15
- GPT-6 Sol and GPT-6 LunaTwo reasoning models with text and image input on both Responses and Chat Completions. For prompts up to 272K tokens, Sol is $2 input, $0.20 cached, $10 output per million tokens; Luna is $0.10, $0.01 and $0.50.
- Sep 22
Two of those entries change what a team must do before it can use a model at all. Astra's restrictions mean a codebase that sets temperature or reads logprobs cannot switch to it by changing a model string, and its tool calling requires the Responses API. The key-governance controls mean the question of who is allowed to create a key now has a setting, which is worth answering before a conference week in which many people will want one. Prices and context limits for the current lineup are kept in our frontier model price index; the launch details for the two new models are in our GPT-6 Sol and Luna post.
03 — The deadlineWhat retires in the days around the event
The deprecations page has three entries that land within a week of DevDay. The first has already happened. The second lands on Monday, the day before the event, which leaves no margin to discover that a batch job still calls a completions-era model.
| Shutdown | What stops | Listed replacement |
|---|---|---|
| September 24, 2026 | Videos API and the sora-2 and sora-2-pro models, including the dated snapshots | None listed |
| September 28, 2026 | gpt-3.5-turbo-instruct, babbage-002, davinci-002, gpt-3.5-turbo-1106 | gpt-5.6-terra |
| October 1, 2026 | GPT-5.4-Cyber | gpt-5.6-cyber |
The four September 28 models are among the oldest of the completions-era lineup, and gpt-3.5-turbo-instruct in particular survives in old scripts written against the legacy completions endpoint. A search across repositories, cron jobs and spreadsheet add-ins for the four strings takes minutes and is the single highest-value item on this list. Our deprecation calendar carries the wider set of dates across vendors.
04 — The workSix checks that fit before Tuesday
None of these depends on what is announced. Each takes an owner and less than a day, and each leaves the account in a state where a new model or feature can be trialled without first cleaning up.
The sessions item deserves the most thought. Our post on the Agents API managed runtime explains what the beta takes over from an application. A team that has built its own compaction and retry logic should know before DevDay whether it wants to keep it, because the announcements most likely to matter to that team will be about this surface, and the comparison is easier to make against a written-down design than against a memory of one.
05 — The standardWhat we will not predict, and why
Every conference produces a week of confident previews built on leaks, listing metadata and inference from job adverts. We wrote down in our post on launch signals how those signals should be read: a model appearing in a catalog is a listing, not a launch, and an unconfirmed name is reported as unconfirmed or not at all. This post follows that rule. It contains no rumoured model, no leaked price and no guess at the keynote.
That is not caution for its own sake. A preparation list that depends on a prediction is only useful if the prediction is right, and a list that depends on the September changelog is useful either way. The six checks above are worth doing if DevDay announces nothing at all.
"DevDay is Tuesday. Before then: grep for the four models that retire Monday, set the key-creation policy and an expiry maximum, list every Chat Completions call that uses tools, and run one costed GPT-Live test. We will read the announcements against that baseline on Wednesday."
06 — AfterWhat to re-check on September 30
The morning after, three pages tell you what actually changed. The changelog, which is where a shipped feature appears with a date and an API path. The deprecations page, because launches often carry a retirement for the model they replace, and the shutdown date matters more than the launch date for existing code. And the pricing page, read against the price index, because a headline price and a per-request bill are different things once cached input, tool calls and per-minute audio are added up.
We keep the running list of releases in our September 2026 release tracker and will update it from those pages, not from the keynote. If your team wants the same discipline applied to its own account, our AI transformation practice runs the inventory in section 04 as a standing service.
07 — ConclusionThe preparation list is the September changelog, not the rumour mill
Clear the four retiring models before Monday, set your key policies, and read Tuesday's announcements against a baseline you already understand
Two working days is enough for six checks. Whatever OpenAI announces on September 29 will be easier to evaluate from an account with no retired models, a key policy and a known voice cost per minute.