ChatGPT Finally Built The AI Employee Claude Couldn't. Launch on this one

Summary

At OpenAI DevDay, the presenter describes a broader shift from AI that waits for individual prompts to AI agents that can work continuously, use software, and take on ongoing responsibilities.

OpenAI: Dots and a shared AI workspace

• Dots are AI workers inside ChatGPT that can keep working after the user closes the app. Each Dot is described as running on GPT‑6 Astra, using its own cloud computer, and connecting to apps through ChatGPT plugins.
• Users can give a Dot ongoing responsibilities rather than one-off tasks—for example, monitoring Teams for new figures and updating a presentation, tracking a software migration, or responding to a canceled supplier order.
• Users can set boundaries and approve important decisions. While a Dot is working, they can check in, ask for a status report, interrupt it with new instructions, and let it continue. Dots are included with ChatGPT Pro; the interviewee says a free version is not planned soon because of cloud-computing and model costs.
• ChatGPT Space provides a shared workspace for users, teams, ChatGPT, and Dots. Its Pages feature lets users work on documents with AI research, charts, images, and data visualizations. The presenter says shared AI-assisted presentations are also expected.
• A more affordable GPT‑6.1 model—rendered as “Soul” in the subtitles—scored 71.4% on OSWorld and 75.2% on DeepSWE. Cached inputs are said to cost $0.10 per million tokens, making extended agent work cheaper.
• Astra Ultrafast is positioned for speed-sensitive tasks, with generation of up to 300 tokens per second. The presenter highlights potential uses in coding, voice assistants, robotics, and games.

Codex, plugins, and automation

• Codex can run programming tasks in the cloud, allowing users to start work from a phone and return later to review the result. A demo involved rewriting a backend in Rust.
• Codex Security Cloud can monitor GitHub repositories and new commits, identify vulnerabilities, and prepare patches.
• Developers can give agents built with the OpenAI Agent API access to control a computer, including opening applications, navigating interfaces, and testing software.
• Other Codex additions include voice control in the command line, an Agent View for managing multiple tasks, and a dedicated code-viewing interface.
• Plugin extensions can display interactive app interfaces—such as editors, calendars, dashboards, and file viewers—inside ChatGPT. ChatGPT can recommend plugins when they become relevant.
• MCP events can trigger work when something changes in another app, while scheduled duties can automate recurring tasks such as weekly project reports. A meeting plugin can take notes, summarize discussions, and send action items to a workspace.
• OpenAI also announced ChatGPT sign-in integrations, shareable developer profiles, and a marketplace for discovering AI products.

The presenter’s main interpretation is that useful agents require more than a capable model: they also need a computer, tools, context, permissions, app connections, and ways to detect changes.

Interview example: AI-assisted software development

In an interview, Thibaut, who works on ChatGPT and Codex, says his team used newer models to combine separate web and desktop ChatGPT codebases. A project estimated to take at least six months was completed in under 30 days, despite involving more than 500 people and, at peak, about 1,000 daily merge requests. The presenter attributes much of the acceleration to the models doing substantial portions of the work.

Other AI announcements covered

• Anthropic Claude Sonnet 5.5: Presented as a faster, less expensive everyday model than the prior Sonnet, with improvements in coding, knowledge work, computer use, and visual understanding. The subtitles claim over 30% faster output and up to 30% lower cost for some tasks. Opus remains the more powerful model, and Haiku 5.5 is mentioned as forthcoming.
• Perplexity Computer automation: Can monitor apps or run scheduled tasks, such as flagging sales problems in Slack or reviewing ad performance each Monday. It can use prior project context and connected services, including Slack, Gmail, Outlook, Linear, and GitHub. The presenter says it uses no credits while waiting for a task.
• Google Gemini 4 Argon: Described as handling long-running tasks, with up to one million initial tokens per run. The video cites use on a codebase exceeding 800,000 lines and a video system reported to run 2.7 times faster after improvements. It is also positioned for research, legal work, long videos, and cybersecurity.
• Perplexity interactive answers: Answers can be explored visually—for example, through interactive stock charts, a bridge simulation, or a model of tidal forces.
• Microsoft MAI Transcribe to Streaming and MAI Voice 2.1: Designed for faster spoken interaction. The transcription model supports 60 languages and can begin recognizing speech in under 100 milliseconds; the voice model supports 23 languages and can respond in about 150 milliseconds.
• Google Health Coach: Uses activity, sleep, recovery, and fitness data to personalize and adjust exercise guidance, such as adapting workouts around an injury, a late workday, or poor recovery.
• AI safety: The presenter reports that major US AI companies, including OpenAI, Anthropic, Meta, Google, and Nvidia, agreed to voluntary safety audits at a White House meeting. Anthropic CEO Dario Amodei is quoted as emphasizing both AI’s potential benefits and its risks.

Reviews, guides, and resources

• The video is primarily a roundup and analysis of product announcements, with demonstrations and an interview; it does not provide a step-by-step tutorial.
• The presenter says timestamps are available for navigating to individual topics.
• A full interview with Thibaut is referenced, covering how he uses Dots, changes to ChatGPT, and his views on AGI. The presenter also mentions a possible future comparison of OpenAI Dots and Meta’s “Mews,” as named in the subtitles.

Main speakers and sources

• Vaibhav Sisinty, the presenter and narrator, reports from OpenAI DevDay and provides the analysis.
• Thibaut, an OpenAI product leader associated with ChatGPT and Codex, is interviewed about Dots and his team’s use of AI in software development.
• The video also references announcements and claims from OpenAI, Anthropic, Perplexity, Google, and Microsoft, as well as statements by Anthropic CEO Dario Amodei.