OpenAI shipped its biggest batch of ChatGPT updates and features in months this July, headlined by the GPT-5.6 model family, a new agent mode called ChatGPT Work, and a redesigned desktop app. Here is what actually changed, what it means for your plan, and what quietly disappeared along the way.
GPT-5.6 and the New Sol, Terra, and Luna Models
GPT-5.6 rolled out globally on July 9, 2026, across ChatGPT, Codex, and the API. Instead of one model, OpenAI split it into three tiers. Sol is the flagship, built for coding, research, and long multi-step work. Terra balances capability and cost for everyday tasks, and Luna is the fastest, cheapest option for high-volume, simple requests.
Where you land depends on your plan. Free and Go users get Terra inside ChatGPT Work and Codex. Plus, Pro, Business, and Enterprise users can choose among Sol, Terra, and Luna and set an effort level for each task. GPT-5.5 Instant still handles fast, everyday chat replies across every tier, and OpenAI has said GPT-5.4 will retire on July 23 now that 5.6 is live.
OpenAI CEO Sam Altman told CNBC that Sol runs about 54 percent more token-efficient on agentic coding tasks than its predecessor, a figure that matters for any team tracking AI spend against output. Early developer reaction has leaned positive, with several builders on X calling GPT-5.6 the most reliable model they have used for daily work so far.
GPT-5.6 also adds two compute modes: max, which gives the model extra room to reason through a hard problem, and ultra, which lets Sol delegate parts of a task to smaller submodels running in parallel. On an internal benchmark for long professional workflows called Agents’ Last Exam, OpenAI reported that Sol scored 53.6.
ChatGPT Work: The New Agentic Mode
ChatGPT Work is the other headline release, and arguably the bigger one for daily use. It is an agent, powered by GPT-5.6, that pulls context from your connected apps and files, then produces finished documents, spreadsheets, presentations, and reports on its own. Unlike a single chat reply, a Work task can run independently for hours while it researches, drafts, and revises.
Work started rolling out first on the Mac and Windows apps for every tier, free included, with the web version to follow.
In practice, that means you can hand ChatGPT Work a rough brief and get back a formatted slide deck, or point it at a shared folder of sales figures and get back a spreadsheet with the analysis already done. The agent can also publish what it builds directly as a Site, so a report or a tracker becomes a shareable page without a separate export step.
A Redesigned Desktop App and ChatGPT Sites
The desktop app changed shape too. Chat, Work, and Codex now live under one switcher in the top-left corner instead of three separate tools. The previous desktop app still exists as ChatGPT Classic and keeps getting model updates and security patches, but new features like Work only ship in the new app. Recents now show chat and Work conversations together, sortable and pinnable, and Work conversations sync across web, mobile, and desktop.
Alongside the redesign, OpenAI introduced ChatGPT Sites, which turns a chat into a working website or lightweight app: a dashboard, a project tracker, a prototype, or an internal portal, previewed and refined inside ChatGPT before you share it. Sites is live now for Business and Enterprise customers.
The new app also adds local file access, so Work can read a document sitting on your computer instead of waiting for an upload. That removes one of the more tedious steps in pulling a report together from scattered sources.
Memory Gets a Quiet Overhaul: Dreaming
The least flashy update might matter the most day to day. OpenAI rebuilt how ChatGPT remembers you under a system it calls Dreaming. Instead of only saving what you explicitly ask it to remember, ChatGPT now runs a background process that reads across your past chats and synthesizes patterns in your preferences, tone, and recurring projects.
Dreaming first arrived in a limited form in April 2025, and the version now expanding to Plus, Pro, and free accounts is more capable and more compute-efficient, according to OpenAI’s own announcement. You can still turn “reference chat history” and “reference saved memories” on or off separately in settings. On Business or Enterprise plans, an admin can isolate memory per project or wipe everyone’s saved memories at once.
OpenAI describes the system around three goals: keeping recent context fresher than old references, carrying a conversational thread even when you return to it weeks later, and picking up details you mentioned in passing rather than only the ones you explicitly flagged. That last point is the real shift. Older memory behaved like a notebook that only filled up when you dictated to it. Dreaming reads the whole conversation and decides on its own what is worth keeping
Pricing, Voice, and What Got Retired
1:Six Plans, with Pro Split in Two
ChatGPT now runs six pricing tiers: Free ($0, with ads in the US), Go ($8 a month), Plus ($20 a month), Pro (either $100 or $200 a month, same models, different usage caps), Business ($20 to $25 per seat), and Enterprise on custom pricing. The $100 Pro tier launched in April 2026 to sit between Plus and the older $200 Pro plan, aimed at users who need heavier Codex or Work usage without paying for the top usage cap.
API access to the GPT-5.5 flagship model is billed separately, at $5 per million input tokens and $30 per million output tokens. Business plans add shared workspaces, SAML single sign-on, and a no-training-on-your-data guarantee, details that matter if procurement is part of the decision.
2:Voice Moves to GPT-Live-1
Paid tiers now run voice conversations on GPT-Live-1, with free accounts on a lighter GPT-Live-1 mini. Both can listen and talk back at the same time, so the exchange feels closer to a real conversation, with natural interruptions instead of the older strict turn-by-turn pattern.
3:What OpenAI Retired
A few features did not survive the year. Canvas, the side panel for writing and code, was folded into the main chat window in late May, replaced by inline writing and code blocks with a full-screen option for longer documents. GPT-4.5 was retired from ChatGPT and custom GPTs on June 26, with existing conversations continuing on GPT-5.5.
Sora, OpenAI’s AI video app, is a bigger casualty. The web and app versions shut down on April 26, 2026, and the developer API follows on September 24. OpenAI’s own account points to plain economics: Sora cost roughly a million dollars a day to run while producing about $2.1 million in total revenue over its lifetime, and usage had already fallen sharply from its launch peak. The Atlas browser is next, scheduled to stop working on August 9, 2026, as its browsing features move directly into ChatGPT and Codex. Anyone still using Atlas should export bookmarks and saved passwords before that date, since browser data does not transfer automatically.
Taken together, the pattern is less about the model getting smarter in the abstract and more about OpenAI moving manual work, drafting, formatting, connecting tools, into the agent layer itself, priced to compete directly with Anthropic’s Claude Max and Claude Cowork.
