Split what is official, what already shipped, and what is still a forecast
- OpenAI has pinned OpenAI DevDay 2026 to Tuesday, 29 September at Fort Mason, San Francisco. The keynote is 10:00 a.m. PT, Sam Altman, livestreamed for free. Event page: devday.openai.com.
- In-person applications are closed; invited registration is $650. Other sessions are recorded for openai.com. Later Exchanges: Bengaluru, Tokyo, Seoul, Paris, Berlin, London, São Paulo, Mexico City.
- GPT-6 Astra is not an unreleased slide. It began rolling out in early September. The API ID is
gpt-6-astra, standard price $10 input / $50 output per million tokens. openai.com/index/gpt-6-astra. - The Agents API entered public beta on 10 September 2026: cloud agents on the hosted Codex harness. No extra platform fee; you pay tokens, tools, and sandbox compute. Introducing the Agents API.
- As of 17 September 2026, OpenAI has not published a DevDay product list. Anything below about what GPT / Codex / API / agents may still add is a model, not news.
“What new products will DevDay ship?” is: what already landed, plus what the keynote may still add. Dating Astra to 29 September is a calendar error.
What you can confirm, and what is not on the invite
Today is 17 September 2026, about twelve days out. The announce post only says “29 September, San Francisco”: Announcing OpenAI DevDay 2026. The event page lists breakfast 8:00, keynote 10:00, lunch 11:30, breakouts 11:15–15:30, closing 16:00, reception 16:45–19:00, all Pacific Time. If you missed the application, watch the livestream. Tickets are not sold at the door.
The room is for engineers, technical founders, and research leads who already build with AI — not a consumer launch. Last year’s DevDay already ran on AgentKit, Codex GA, and the Apps SDK. In the two weeks before this show, OpenAI put GPT-6 Astra and the Agents API in developers’ hands. 29 September looks like an upgrade-and-close day, not the first appearance of a name that does not exist yet.
| Item | Status | Evidence |
|---|---|---|
| Date and venue | Confirmed | Official: 29 September, Fort Mason |
| Keynote / livestream | Confirmed | Official: 10:00 PT, free |
| In-person ticket | Applications closed | Official: $650 if invited |
| GPT-6 Astra | Live before the show | Official: paid ChatGPT + API |
| Agents API | Public beta | Official: 10 September post |
| Day-of product list | None | No official keynote SKU sheet |
| GPT-7 / a new brand name | None | Any specific new name is a scenario |
Official versus model stays labeled. This title does not say “hands-on,” and we are not in Fort Mason. Prices are from OpenAI docs and can still move before the show.
GPT: Astra is already callable — what DevDay may still add
If you searched “DevDay GPT,” do not treat GPT-6 Astra as a 29 September gate. The model page calls it the strongest end-to-end work model for hard reasoning, coding, computer use, research, and long documents. reasoning.effort goes low / medium / high / xhigh / max. Context is about 1.05 million tokens, max output 128k, knowledge cutoff 30 April 2026. GPT-6 Astra Model. How it sits next to 5.6: GPT-6 Astra vs GPT-5.6.
| Model (official now) | Input / 1M | Output / 1M | Notes |
|---|---|---|---|
gpt-6-astra | $10 | $50 | Cache read $1 / write $12.50; long context >272K costs more |
gpt-5.6-sol | $4 | $20 | Short-context row; API pricing |
gpt-5.6-terra | $2 | $12 | Everyday coding tier |
gpt-5.6-luna | $0.20 | $1.20 | Cheap classify / extract |
gpt-5-pro | $15 | $120 | Responses API only; GPT-5 Pro |
Astra also has Fast mode: up to about 2× speed at 2× standard price. Batch / Flex are 50% of standard. On ChatGPT it is rolling to Plus / Pro / Business / Enterprise; Free and Go do not get it. Enterprise workspaces ship off; an admin must enable it. Pro / Business / Enterprise also get GPT-6 Astra Pro — that SKU is already in the launch post, not a forecast.
Model forecast: the keynote is more likely to do one or two of these — clarify Astra Pro quotas or regions; add a cheaper “fast / mini” SKU so not every call hits Astra; or spell out default models and fees for computer use and tools. There is no public signal for a “GPT-7” name. Price tables often move in event week; trust developers.openai.com/api/docs/pricing on the day.
Codex: already a platform — DevDay is about workflow, not a second invention
Codex went generally available at DevDay 2025, with an SDK and Slack-class entry points. On 19 August 2026 OpenAI framed Codex as an open agent harness: CLI, app-server, and official SDK are open source. Scripts use codex exec; products that need durable threads and approvals use app-server. Codex as a platform. How it shows up in ChatGPT Work: ChatGPT Work and Codex.
The other pre-show line is already wired: the Agents API runs that same Codex harness, with OpenAI hosting orchestration and session state. If you want the loop in your process, use the Agents SDK or talk to the Responses API yourself. Official map: Agents guide. Do not hear “Codex the product” and “Agents API” as rival brands — one is the coding agent you open; the other is the hosted interface to the same harness.
Forecast: the show is more likely to add enterprise controls, deeper CI / IDE hooks, usage and audit, or a clearer hosted-sandbox quota card. Hardware side quests (Codex Micro and the like) are not the event-page spine; do not treat them as a 29 September must-ship. How another open harness splits the job: What DeepSeek Harness is.
API and Agents API: the beta fence matters more than the slogan
The official Agents API examples create a production-shaped agent in one call: task, model, tools, environment. The environment can be an OpenAI-hosted sandbox, your own machines, or a partner sandbox. Docs list automatic context compaction, multi-agent, programmatic tool calling, and MCP. Limits are also on the page: US data residency only, no Zero Data Retention; a self-hosted sandbox does not make the API ZDR-eligible. Agents API overview.
Billing: the API itself has no extra “Agents platform fee” in beta. You still pay model tokens, tool calls, and hosted containers. That is not “free” — a long Astra run plus computer use still follows the rate card. The Responses API remains the low-level loop you control; Chat Completions still exists, but new work defaults toward Responses / Agents.
Forecast: the API sentences worth hearing on 29 September are whether beta ends, whether residency grows, whether ZDR gets a door, and which sandbox partners show up. Those decide production, not another slide that says agents matter. Rolling your own three layers: the 2026 AI agent trio.
How to prepare: a watch list, not a bet on a product name
In twelve days you can pin live billing and integrations. You cannot inventory a SKU that does not exist.
| Listen for | Official now | Model (not news) |
|---|---|---|
| A new model ID | Astra already in the catalog | More likely a tier / quota, not GPT-7 |
| Agents API GA or not | Public beta + US residency | GA, or more regions / partner sandboxes |
| Codex billing and limits | CLI / SDK / app-server already open | Enterprise audit, CI, quota card |
| Tool / computer-use unit prices | Pricing page lists tool fees | Default model or discount may move |
| What stays in preview | Agents API is explicitly beta | Hear what still cannot go to prod |
Practice: pin Astra requests to a snapshot; keep everyday coding on 5.6-sol / terra; if you need compliance, do not pour production into an Agents API that still has no ZDR. If you need a Mac that does not sleep through a sandbox and MCP, use the rent-a-Mac onboarding checklist — a laptop lid is not a 24-hour agent host.
FAQ
Will OpenAI DevDay 2026 launch GPT-7?
There is no public signal. The main new model before the show is already GPT-6 Astra. Putting “GPT-7” in a headline is a scenario, not the official schedule.
Can I use the Agents API now, or do I wait for DevDay?
You can use it now. It is public beta, announced 10 September. Do not wait until 29 September for a first call. Production still depends on residency, ZDR, and limits.
Is GPT-6 Astra locked until DevDay?
No. Paid ChatGPT tiers and the API are already rolling out. Enterprise is off by default. Free and Go do not get it.
Are Codex and the Agents API the same thing?
Not the same product door, but they share the Codex harness. Codex is the coding-agent product (CLI / IDE / ChatGPT). The Agents API hosts that harness. The Agents SDK runs the loop in your process.
Can I still buy a ticket? Does the livestream cost money?
Applications are closed; no door sales. The keynote livestream is free. Other sessions land on openai.com after the day.
What to watch next
Which new products DevDay 2026 will add is, today, a calendar plus a stack that already shipped: 29 September at Fort Mason, 10:00 PT livestream; Astra on the API; Agents API in public beta. Names that are not on an official page are forecasts.
On keynote day, check three lines: a new model ID or not, whether Agents API is still beta, and whether computer-use or long-context prices moved. Until then, pin routing and bills to the docs, not to a product name on a social feed.
A long agent task cannot follow a laptop lid
Agents API and Codex pay off when the job finishes: sandbox, MCP, multi-hour sessions. Sleep, conference Wi-Fi, and a closed lid kill that experiment first. Unified memory, low idle watts, and a native macOS Unix toolchain already run on cloud Mac mini hosts you can rent. Burn down Astra versus 5.6-sol bills and isolated sandbox dirs on a daily lease before you cut traffic after 29 September.