ChatGPT Pro now fans one task out to several agents, and the Agents API stopped charging for idle sandboxes
OpenAI posted at 20:00 UTC on October 16 that GPT-6 Pro in ChatGPT got an upgrade, then edited the model name to GPT-6 Astra Pro in a repost 27 minutes later that went on to 1.98 million impressions and 11,300 likes. The upgrade is faster responses, better use of plugins and tools, and the new part: multiple agents can tackle different parts of a task in parallel, you can follow each agent's progress, and open completed tasks to review their work. It is live for Pro users in the Chat tab, with availability in managed workspaces depending on workspace settings. The release notes say the same in three bullets and nothing more. There is no openai.com post, no number of agents, no word on whether parallel agents consume Pro message limits per agent, which was the first complaint in the replies.
That makes it the first multi-agent fan-out inside ChatGPT proper, as opposed to Codex or Dots, and it arrives at the consumer tier with a review UI attached: progress per agent, a completed-tasks view, which is OpenAI's answer to the question every parallel-agent product gets, who checks the five pieces of work five agents produced. The replies asked that question within the hour. The same evening OpenAI showed the other end of the stack: an OpenAI engineer's dot triaged a 3am on-call incident on its own, investigated, then opened a Codex thread the engineer steered from the ChatGPT app. The Codex team also published how OpenAI engineers use Codex to dig through the codebase and chip away at the maintenance backlog.
For builders the more concrete news came at 02:05 UTC on October 17 from the OpenAI developers account's weekly Agents API note. GPT-6 Astra in Ultrafast mode is now available in the Agents API. Sandbox snapshots let you restore an OpenAI-hosted environment after idle with files intact and keep the work going. And billing changed: you no longer pay for environments that are idle. The post was truncated at an item about environment keys and there is no changelog entry yet, so take the details as a tweet. The idle-billing change is the one to notice. Hosted agent runtimes have been quietly charging for persistent state while an agent waits on a human, which is the denial-of-wallet shape we wrote about in September and, as of this week, the shape of most agent sessions: Anthropic's own employment report says its interactive sessions sit idle about 80 percent of the time waiting for people. Snapshots plus no idle charge is the product form of that number.
Links: https://x.com/OpenAI/status/2111192259073094035 , https://help.openai.com/en/articles/6825453-chatgpt-release-notes , https://x.com/OpenAIDevs/status/2111277450688892969 , https://x.com/OpenAI/status/2111215293737599284 , https://x.com/OpenAIDevs/status/2111213093577691337
← Back to all articles
That makes it the first multi-agent fan-out inside ChatGPT proper, as opposed to Codex or Dots, and it arrives at the consumer tier with a review UI attached: progress per agent, a completed-tasks view, which is OpenAI's answer to the question every parallel-agent product gets, who checks the five pieces of work five agents produced. The replies asked that question within the hour. The same evening OpenAI showed the other end of the stack: an OpenAI engineer's dot triaged a 3am on-call incident on its own, investigated, then opened a Codex thread the engineer steered from the ChatGPT app. The Codex team also published how OpenAI engineers use Codex to dig through the codebase and chip away at the maintenance backlog.
For builders the more concrete news came at 02:05 UTC on October 17 from the OpenAI developers account's weekly Agents API note. GPT-6 Astra in Ultrafast mode is now available in the Agents API. Sandbox snapshots let you restore an OpenAI-hosted environment after idle with files intact and keep the work going. And billing changed: you no longer pay for environments that are idle. The post was truncated at an item about environment keys and there is no changelog entry yet, so take the details as a tweet. The idle-billing change is the one to notice. Hosted agent runtimes have been quietly charging for persistent state while an agent waits on a human, which is the denial-of-wallet shape we wrote about in September and, as of this week, the shape of most agent sessions: Anthropic's own employment report says its interactive sessions sit idle about 80 percent of the time waiting for people. Snapshots plus no idle charge is the product form of that number.
Links: https://x.com/OpenAI/status/2111192259073094035 , https://help.openai.com/en/articles/6825453-chatgpt-release-notes , https://x.com/OpenAIDevs/status/2111277450688892969 , https://x.com/OpenAI/status/2111215293737599284 , https://x.com/OpenAIDevs/status/2111213093577691337
Comments