OpenAI’s event page put DevDay 2026 on September 29 in San Francisco, and its own recap counts more than 20 major announcements. For people who call the API, the items with dated changelog entries and docs behind them are the Agents API (public beta since Sep 10, computer use added Sep 29), GPT-6.1 Sol, Ultrafast mode and the Decisions API beta.
I read every page below on 2026-10-14 and have not run any of it myself. This recap separates what is shipped and documented from what is preview or only announced, links a primary source for each item, and ends with a plan for this week.
1. Context: Three Weeks of Changelog, One Event
DevDay did not arrive alone. The API changelog shows the pieces landing in order: the Agents API beta on Sep 10, GPT-6 Sol and Luna on Sep 22, the DevDay-day entries on Sep 29 and follow-ups through Oct 8. I use that changelog as the spine because every row has a date and a docs link, which the recap page does not.
What the API changelog dates (2026)
How I label things
Shipped and documented means a dated changelog entry or a docs page I could read. Preview means the vendor itself calls it preview or beta. Announced means it appears in the recap but I found no docs behind it. Where OpenAI’s pages say nothing, I say nothing.
Already covered here: GPT-6 Sol and Luna pricing · the Oct 23 deprecations checklist
One correction to what circulates: Ultrafast is not a separate model. It is a service tier you select with service_tier, at six times the standard price, on models you may already call.
2. Four Implications for Developers
These are the four changes that can touch an architecture or a bill this quarter, ordered by how much code they move.
A managed agent loop you do not run
Agents API · public beta since Sep 10
OpenAI runs sessions, orchestration, context compaction and recovery; you supply tools and choose where commands execute. The docs say usage bills at the model’s API rates, that data residency is US only and that Zero Data Retention is not supported. Check those lines before anything else.
Computer use with your app in the loop
Agents API · Sep 29
The agent drives an OpenAI-hosted browser, and your app must answer each new-origin approval and any sign-in request. Passkeys and QR sign-in are not supported. Approving an origin does not confirm individual actions, so purchases and destructive steps need a runtime you control.
Ultrafast is a pricing decision
service_tier: ultrafast · 6x standard
Available for gpt-6-astra since Sep 29 and gpt-6.1-sol since Oct 8. Sol goes from $2/$10 to $12/$60 per 1M input/output tokens. OpenAI’s Codex docs say up to 8x faster than Astra standard in Codex; the API guide publishes no tokens-per-second figure, so measure your own prompts.
The work moves off your laptop
Codex cloud · Bedrock Managed Agents
Codex cloud runs tasks in reusable environments from any device, and Codex Security Cloud (research preview) scans repos once or continuously. Bedrock Managed Agents moves the harness into AWS with IAM SigV4 auth; AWS labels it Preview and OpenAI warns the API contracts are not identical.
The two calls worth trying first
Adapted from OpenAI’s Agents API quickstart and Ultrafast guide, shortened. I have not run it, so read it as the documented shape, not a tested recipe:
from openai import OpenAI
# 1) Agents API session (key scopes: api.agents.read, api.agents.write, api.responses.write)
with OpenAI() as client:
with client.beta.agents.sessions.create(
agent={
"model": "gpt-6-astra",
"instructions": "Write clean code, run it, and report the actual output.",
},
environment={"type": "openai_hosted"},
input="Create tree.py that prints the file tree of this directory, run it, show the output.",
stream=True,
).with_result_collection() as stream:
for event in stream:
print(event.to_json(indent=None), flush=True)
result = stream.get_final_result()
print(result.output_text, result.session_id)
# 2) Ultrafast: a service tier on the Responses API (gpt-6-astra or gpt-6.1-sol)
client = OpenAI()
response = client.responses.create(
model="gpt-6.1-sol",
service_tier="ultrafast",
input="Summarise this diff in two sentences.",
)
print(response.output_text)Hosted does not mean supervised
Origin approval gates where the browser may go, not what it does there. OpenAI’s docs tell you to treat website content as untrusted and to keep screenshots and sign-in values out of logs. Build your approvals UI as if a hostile page were reading over the agent’s shoulder.
3. What to Do This Week
Do not adopt anything on a Tuesday because a keynote said so. This is the five-day sequence I would run with a team that already ships on OpenAI:
☀️ Days 1–2: Check the blockers
- 1. Read the data lines first: Agents API residency is US only with no Zero Data Retention; Ultrafast on Astra has no EU option, while Sol Ultrafast supports US and EU.
- 2. List which workloads are long-running loops you hand-built (Agents API candidates) and which are latency-bound (Ultrafast candidates). Ignore the rest.
- 3. Have an admin review key rules: project keys can now expire (Sep 10) and key creation can be restricted per org or project (Sep 15).
📅 Days 3–4: Pilot behind a throwaway key
- Agents API: mint a key with api.agents.read, api.agents.write and api.responses.write, keep it outside the sandbox, and run the quickstart task. Count success only on agent.session.turn.completed; idle alone is not success.
- Ultrafast: replay one eval set on standard and ultrafast over a WebSocket, then compare time between tokens and cost at 6x, not just pass rate.
- Computer use: log every origin approval, deny by default, and never route credentials through ordinary messages.
📆 Day 5: Budget and calendar
- Cost: price long prompts too: above 272K input tokens Sol bills 2x input and 1.5x output for the whole request.
- Evals: rerun image-input evals on GPT-6 Sol and Luna, because OpenAI fixed an image-encoding bug on Sep 25.
- Calendar: add the Oct 23 shutdowns, the Decisions API GA (its page says “in the coming weeks”) and today’s GPT-5.5 retirement in Codex for ChatGPT sign-ins (API-key sessions are unaffected).
4. Shipped, Preview, or Only Announced
This is the ledger I would hand a team lead. Each row says what I could verify and where; the red box lists what I could not.
Shipped and documented: API
Agents API (public beta), computer use in the Agents API, GPT-6.1 Sol ($2/$10, 1,050,000-token context), Ultrafast mode, and the Decisions API (public beta, gpt-6-luna only, $0.10 per 1M input tokens, GA expected in the coming weeks).
Shipped and documented: Codex
Codex cloud environments (the recap lists Plus, Pro, Business, Healthcare, Education and Enterprise), code review for GitLab merge requests (beta), and GPT-6.1 Sol as the CLI default since 0.161.0.
Preview
Codex Security Cloud is a research preview that needs workspace access and a connected GitHub repo. Bedrock Managed Agents is labelled Preview on AWS’s page, with no regions or pricing listed; OpenAI’s own docs defer to AWS for both.
Announced, no docs I could verify
Zero Data Retention with Private Safety Processing as a project policy appears in the recap only as an image caption. Dots, ChatGPT Space, plugins and the OpenAI Marketplace are ChatGPT-side items, not API surfaces, so I left them out.
What I could not confirm
A figure of roughly 300 tokens per second for Ultrafast is circulating; none of OpenAI’s pages I read state an absolute number, only “up to 8x faster” in Codex and a 6x price. One secondary site gives Sep 11 for the Agents API beta, while the changelog says Sep 10. A third-party blog claims a Bedrock public preview in three US regions on Sep 29; AWS’s page names no regions.
Golden rule for this quarter
Read the data-residency and retention lines before the feature list, price Ultrafast at 6x before you route traffic to it, and pilot every agent surface behind a throwaway key.
Conclusion
DevDay 2026 was less about one big model and more about OpenAI moving the agent loop, the sandbox and the speed dial into the platform. The documented parts are real and dated; the rest is a recap page with pictures.
Run the week above and you will know, with your own numbers, which of these belongs in your stack and which was just a good keynote.
This week in one screen
Verify
- • Residency and retention
- • Key scopes and expiry
- • Model ids in production
Pilot
- • Agents API session
- • Ultrafast A/B at 6x
- • Logged origin approvals
Watch
- • Decisions API GA
- • Bedrock preview terms
- • Oct 23 shutdowns
Sources
- OpenAI — Announcing OpenAI DevDay 2026 (event page)
- OpenAI — DevDay 2026 Recap
- OpenAI — API changelog (Sep 1 to Oct 8, 2026 entries)
- OpenAI — Agents API overview and quickstart
- OpenAI — Computer use in the Agents API
- OpenAI — Ultrafast mode guide
- OpenAI — GPT-6.1 Sol model page
- OpenAI — Decisions API guide
- OpenAI — Bedrock Managed Agents (docs)
- AWS — Amazon Bedrock Managed Agents, powered by OpenAI (Preview)
- OpenAI — Codex changelog
- OpenAI — Codex Security overview



