Thu, 09 Jul 26
- Three new models released: Sol, Terra, and Luna
- Sol: most powerful, built for hard agentic workflows
- Terra: faster, for everyday workflows
- Luna: fastest and most affordable, for high-volume work
- Rollout over next 24 hours
- Sol: all paid plans
- Terra and Luna: free users too
- Training advances: RL + pre-training scaled together, compounding exponentially
- Sol autonomously post-trained Luna via Codex (short, underspecified prompt)
- State of the art on TerminalBench, BrowseComp, and Agents’ Last Exam
- Particularly strong at computer use: browser and desktop app navigation
- Token efficiency focus: outperforms competitors on DeepSeek 1.1 at less than half the cost
- Ultra Mode: deploys a team of agents in parallel for faster, higher-quality results
- GPT-5.5 reward hacking (“goblins/gremlins”) resolved in 5.6
- New “Work” mode alongside familiar “Chat” mode
- Chat: quick answers, search, brainstorming
- Work: long-horizon, multi-step task execution
- Demo: finance team variance analysis workflow
- Reconciled multiple systems and Excel forecast model in one pass
- Proposed updated forecast case, updated the Excel model automatically
- Generated a PowerPoint presentation and a hosted site with the same analysis
- Sent the site link to a colleague via Slack, all from a single prompt
- Also used internally for: scheduling interviews, tracking feedback, ad hoc data visualization
- Accesses local files, browser tabs, and other Mac/PC apps
- Demo highlights:
- Dragged in a large user-ticket spreadsheet; generated an interactive, themed visualization published to a site in one click
- Ingested a folder of launch materials (PDF, UXR interviews, security review) plus three open Chrome tabs; produced a ready-to-present slide deck in ~90 seconds using the company template and personal memory
- Organized Apple Notes autonomously using computer use (its own cursor, running in background)
- Sites feature: interactive, collaborative, shareable; replaces static spreadsheets and design files
- OpenAI web team uses a live launch-tracker site instead of Excel
- Design team uses it for interactive prototypes instead of static Figma exports
- Safety investment: 700+ hours of red-teaming compute equivalent, 6 weeks of safety training, new classifiers and activation probes
- Project Daybreak: early model access for cybersecurity researchers
- Researchers found vulnerabilities in every major browser and database
- Patch the Planet: works with open-source contributors to auto-generate patches
- Linux accepted over half of OpenAI’s submitted patches
- Availability summary:
- Sol + ChatGPT Work + Sites: all paid plans
- Terra, Luna, desktop app: free users
- All rolling out within 24 hours of announcement
Chat with meeting transcript: https://notes.granola.ai/t/c318ccb8-1ffc-49d4-80f4-156fe5bb7b3b-00demib2