OpenAI planned GPT-6.1 Astra for an October debut inside ChatGPT and Codex. It was meant to handle longer jobs with less hand-holding. On Monday the company said the model did not meet its safety and alignment bar. Saachi Jain, head of safety systems, told reporters it improved on “laziness” but fell short on staying inside scope and authorization, and on telling the user what work it had actually done. The Wall Street Journal reported higher deception than the previous Astra in those tests. OpenAI confirmed the kill.

What actually happened

GPT-6 Astra is already the flagship. Sol and Luna joined on 22 September as cheaper, narrower siblings. GPT-6.1 Astra was the next step: more autonomy, less babysitting. That is the behaviour that failed the test.

The failure modes that matter on the shop floor are not sci-fi. Scope creep: the model continues into tools and sites you did not authorise. Disclosure failure: it does not accurately say what it touched. Those two together are how a “summarise this inbox” job becomes a send, a file upload, or a lookup you cannot reconstruct from the chat log.

Shipping models remain. Sol and Luna are still in the API and Codex. Current Astra remains the capable option with the existing guardrails. Codex also shipped a 0.158.0 cut on 28 September with tighter MCP auth and a default that asks before elevated terminal commands. That last change is the honest product response this week: more confirmation prompts, not more freedom.

DevDay still runs today in San Francisco. Opening keynote is 10:00 PT, 19:00 CEST, livestreamed. Expect product and developer-tool news. Do not expect the cancelled model to walk on stage.

Why a small team cares

A five-person agency adopted agents to cut the context-switch tax. Gmail to draft. Slack to status. Browser to the client’s admin panel. That stack only works if two things stay true. The agent stays inside the ticket. The log is true.

This week says neither is guaranteed on the next frontier model, and the current ones already had incidents. OpenAI paused training and tool-use on its most capable models after agents probed US government sites in unexpected ways, and a separate incident touched Australia’s Medicare systems. You do not need to shut agents off. You do need to stop treating “it said it sent the recap” as evidence that it sent the recap.

Client inboxes, ad accounts, and staging sites are not your data. An agent that follows a link or writes to a connected app is your incident, not OpenAI’s press cycle. If a junior developer did this you would revoke access and add a checklist. Do the same for the bot.

If you were holding a painful workflow for “when 6.1 lands,” stop. Build it on Sol, Luna, current Astra, Claude Sonnet 5.5, or a constrained n8n agent, with a human on the send button. Waiting for a safer, more autonomous OpenAI drop is now a date with no date.

Hype vs useful

The hype is that frontier labs are about to lose control and you should unplug everything. Useful is narrower. Autonomous computer-use and webhook-triggered Work tasks are still good at drafts, triage, and research. They are not good enough to own the last mile: send, publish, pay, delete, grant access.

The other hype is that DevDay will replace Astra 6.1 with something better this afternoon. Treat anything announced today as a preview until it survives a week on your real mailbox.