OpenAI spent a single week in July recasting ChatGPT from a chatbot into what it calls a 'colleague,' shipping three products at once: the GPT-5.6 model family, ChatGPT Work, and a web-builder called Sites. The through-line is a shift from answering questions to delivering finished work -- an agent that returns a completed slide deck or a populated spreadsheet, not a list of bullet points telling you how to make one.
Key facts
- GPT-5.6 family: three tiers -- Sol (flagship), Terra (balanced, GPT-5.5-level at lower cost), and Luna (budget).
- API pricing: Sol $5/$30 per million input/output tokens, Terra $2.5/$15, Luna $1/$6.
- ChatGPT Work: an autonomous agent for multi-hour and multi-day projects, folding the old Codex app into a unified Chat/Work/Codex interface.
- Primary source: OpenAI's announcements index.
Start with the model family, since it is the engine underneath. GPT-5.6 comes in three named tiers. Sol is the flagship, tuned for coding, cybersecurity, science, and heavy knowledge work. Terra is the sensible middle -- roughly GPT-5.5-level intelligence at a materially lower price. Luna is the cheap, fast option for high-volume everyday tasks. OpenAI is also pushing a claimed leap in 'design judgment,' arguing GPT-5.6 can build interfaces that are 'elegant, intuitive, and actually functional' rather than the generic layouts earlier models produced -- a bet that visual and product taste is now a differentiator, not just raw reasoning.
The centerpiece is ChatGPT Work, and it is the clearest expression of the 'colleague' framing. Where a chatbot answers a turn and waits, ChatGPT Work takes a goal, searches your connected apps -- CRM, spreadsheets, documents -- breaks the goal into sub-tasks, and executes them over hours or days. It has a Plan Mode for reviewing the approach before it runs, and scheduling, so you can tell it to 'run this report every Monday morning.' Critically, it delivers artifacts: a finished spreadsheet, a slide deck, a document, or a working web app. OpenAI also folded its standalone Codex coding app into this same interface, so inline diff editing, pull-request reviews, and multi-repo work now live inside one Chat/Work/Codex surface. Think of it less as a smarter search box and more as a junior teammate you delegate a project to and check on later.
Sites is the exit ramp. It is a no-code, chat-driven builder for dashboards, reports, and simple interactive web apps -- and it plugs into the end of a ChatGPT Work run. An agent can gather data, analyze it, and then 'publish' the result as a shareable Site, turning an analysis into something a colleague can actually open in a browser. Together the three pieces describe a loop: a frontier model does the thinking, an agent does the multi-step doing, and Sites does the shipping.
Why it matters is competitive positioning as much as capability. This is OpenAI planting a flag in autonomous agents and knowledge-work delivery just as Anthropic, Google, and open-weight challengers crowd the frontier-model tier. The pitch to enterprises is no longer 'a better assistant' but 'a worker that finishes things.' The honest caveat: multi-day autonomous agents are exactly where reliability gets hard -- long-horizon plans drift, a wrong sub-task compounds, and 'delivers a finished deck' is only useful if the deck is correct. OpenAI is selling the vision of a colleague; whether it behaves like a reliable one across real multi-day projects is the thing that will actually be tested in the months after launch, not in the demo.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)