OpenAI Launches GPT-5.6 and ChatGPT Work, Its Workplace Agent

On Thursday, OpenAI publicly released its GPT-5.6 family of models and pushed a new workplace agent called ChatGPT Work into paying accounts. The launch lands roughly two weeks after OpenAI held GPT-5.6 to a limited partner preview at the U.S. government’s request, and in the same week Meta rolled out Muse Image, an in-house image generator from its Meta Superintelligence Labs division. OpenAI also turned on a new generation of voice models, GPT-Live, that listen and speak at the same time, joining the package at a moment when more than 150 million people already talk to ChatGPT each week. The new ChatGPT desktop app for macOS and Windows shipped the same day, putting Chat, ChatGPT Work, and Codex into a single window.

ChatGPT Work pulls context from the apps and files a user has already connected and produces finished documents, spreadsheets, presentations, reports, and websites in response. The agent surfaces clarifying questions mid-task, changes course when the brief shifts, and pauses for approval at consequential moments. At launch, Work is rolling out to Pro, Pro Lite, Enterprise, and Edu users on web and mobile, with Plus and Business plans following over the coming days, per OpenAI’s release notes. Anthropic launched its own workplace agent, Claude Cowork, in January, and Google has been threading Gemini deeper into Workspace through 2026. Three companies are now selling similar finished-deliverable assistants that run inside the apps their target users already have connected.

What ChatGPT Work Actually Does

OpenAI’s release notes describing ChatGPT Work call it “an agent for longer, more involved tasks.” It pulls context from connected apps and files, returns finished work, and keeps the user in the loop at consequential steps. The mode sits in the same product as OpenAI’s coding agent Codex and behaves like a co-pilot that returns finished work for review.

Users give Work a multi-step goal at the start, then watch the agent research, plan, execute, and return with a finished deliverable. The agent asks clarifying questions mid-task and prompts for approval before consequential actions such as sending a message or replacing a file. OpenAI is folding the separate ChatGPT Scheduled Tasks launch into Work, so jobs can fire on a calendar, on a trigger, or as monitors that watch for changes. ChatGPT Sites, the feature that turns a Work session into a publishable interactive website or lightweight app, is exposed inside Work as well. Work also includes Plugins, the new packaging format that replaces OpenAI’s old App Directory.

OpenAI’s bundled task list for ChatGPT Work covers:

• Research and analyze information across connected apps and files
• Create finished documents, spreadsheets, presentations, reports, and Sites
• Run multi-step workflows that pull context from apps the user already has connected
• Fire Scheduled Tasks once, on a schedule or trigger, or as monitors that watch for changes
• Ask clarifying questions and surface approval prompts at consequential moments

On the rollout, Pro, Pro Lite, Enterprise, and Edu users receive Work first on web and mobile, and Plus and Business plans follow over the coming days. Enterprise and Edu workspaces start with a two-week preview period in which Work is off by default and admins can opt out before it turns on automatically. Work is available in every region ChatGPT itself supports, including the EEA, Switzerland, the United Kingdom, and the UAE. The public release ends two weeks of government-coordinated staging that began when GPT-5.6 first hit a small set of trusted partners on June 26, per earlier reporting on the partner-only preview.

Three Models, Three Prices

GPT-5.6 is not a single model. The GPT-5.6 family preview page introduces a family of three: Sol, the flagship aimed at complex reasoning and programming; Terra, the balanced option for everyday work; and Luna, the fastest and most affordable. Sol sets what OpenAI calls a new state of the art on Terminal-Bench 2.1, a benchmark for command-line workflows that test planning, iteration, and tool coordination. On GeneBench v1, which tests long-horizon genomics and quantitative-biology analysis, Sol scores higher than GPT-5.5 while using fewer tokens. For cybersecurity on ExploitBench, Sol is competitive with Anthropic’s Mythos Preview while consuming roughly one-third as many output tokens.

The GPT-5.6 family is priced per million tokens: Sol at $5 input and $30 output, Terra at $2.50 and $15, and Luna at $1 and $6. OpenAI positions Terra as competitive with GPT-5.5 at a lower per-token cost, and Luna as the lowest-cost tier. ChatGPT Pro adds a higher-capability variant, GPT-5.6 Sol Pro, which powers the Pro reasoning mode on Pro accounts. Inside standard ChatGPT, GPT-5.6 Sol drives the Medium, High, and Extra High reasoning levels on eligible paid plans. GPT-5.5 Instant remains the default for fast responses. GPT-5.6 Terra and Luna live in Work and Codex and through the API rather than in standard ChatGPT chats.

The naming itself signals the new shape of the family. OpenAI says the number identifies a generation while Sol, Terra, and Luna mark durable capability tiers that will advance on their own schedule. That split gives developers and ChatGPT users cleaner choices across intelligence, speed, and cost. Each tier carries its own price, its own availability, and its own rules for which OpenAI surface reaches it.

GPT-5.6 family at a glance

The Two-Week Pause at the Government’s Request

GPT-5.6 was supposed to land broadly on June 26, the day OpenAI first showed the family. Instead, at the U.S. government’s request, OpenAI held the public release to a small group of trusted partners whose participation had been shared with the government. A few weeks earlier, Anthropic had been forced to disable global access to Claude Fable 5 and Mythos 5 after the Commerce Department issued an export control directive; that directive was lifted in late June, after which Anthropic restored access. The Trump administration’s June 2026 executive order on frontier AI set up a voluntary pre-release review process. OpenAI worked through the process, ran the safety testing, and on Thursday flipped the switch for everyone.

Sam Altman told CNBC the company made “many changes” through a “collaborative back and forth” with the administration, and called the government’s technical capabilities “impressive.” OpenAI’s GPT-5.6 launch ships with protections trained into the model, real-time misuse classifiers that can pause generation mid-stream, and account-level monitoring. The company dedicated 700,000 A100-equivalent GPU hours to automated red-teaming aimed at universal jailbreaks. OpenAI says it does not want the partner-only preview to become the long-term default, calling it a “short-term step” on the path to broader availability.

Key dates around the rollout

• OpenAI first previewed GPT-5.6 on June 26, 2026
• Initial rollout limited to a small group of trusted partners shared with the government
• GPT-Live shipped on July 8, 2026
• GPT-5.6 public release on July 9, 2026
• Anthropic Claude Fable 5 and Mythos 5 export controls lifted in late June 2026

Anthropic, Google, and Meta Are Already Here

ChatGPT Work did not appear in a vacuum. Anthropic unveiled Claude Cowork on January 12, 2026. Google has been threading Gemini deeper into Workspace through the first half of the year, with documented updates to Gemini in Sheets, Docs, Slides, and Drive and a Cloud Next ’26 push toward agent-first enterprise tools. Meta released Muse Image on July 7 as an in-house image generator. It is the second major release from Meta Superintelligence Labs, led by Alexandr Wang, after the April unveiling of the Muse Spark large language model that succeeded the company’s previous Llama family. The contest is over which software takes the meeting prep, the deck draft, the CRM update before the user opens the next tab.

Meta positions Muse Image as an advertiser tool for spinning up on-brand variants with fewer iterations and as a way off third parties like Midjourney and Black Forest Labs. The Muse image model pairs with Muse Spark, the language model Meta unveiled in April to replace its Llama family. Internal benchmark tests show Muse Image trailing OpenAI’s GPT Image 2 and beating Google’s Nano Banana 2 in image-editing tasks. Meta plans to release Muse Video at a later date.

Each of the four companies frames its agent around a different theory of the office. Anthropic’s Cowork positions the model as an autonomous teammate that ships finished deliverables from a goal the user hands it. OpenAI’s Work promises finished documents, with Scheduled Tasks and admin-controlled rollout as the connective tissue. Google is rolling Gemini into the apps enterprises already pay for, lowering the swap-in cost. Meta is skipping the office entirely and going straight after the surface where advertisers spend.

The pricing structure tells a parallel story. OpenAI is selling GPT-5.6 by the token to developers while selling ChatGPT Work to enterprises by the seat. Anthropic is matching that logic with Claude Cowork, where enterprise plans bundle the Claude usage needed to run the agent. Google is selling Workspace access plus a paid Gemini upgrade, layered onto existing seat subscriptions. The four are converging on the same product shape: an agent layered on a frontier model with the workplace software stack beneath it.

The office AI agents already in market

1. OpenAI ChatGPT Work: launched July 9, 2026; available on web, mobile, and the new desktop app; powered by GPT-5.6 Sol, with Terra and Luna in Work and Codex.
2. Meta Muse Image: released July 7, 2026 inside Meta AI, WhatsApp, and Instagram Stories; positioned for advertisers and creators rather than office workers.
3. Google Gemini in Workspace: expanded across Docs, Sheets, Slides, and Drive through 2026; pitched as an agent-first enterprise platform at Cloud Next ’26.
4. Anthropic Claude Cowork: launched January 12, 2026; runs as an autonomous agent on local files and apps; plugin support added January 30.

Voice and the Rest of the Day-One Bundle

OpenAI shipped a new voice system on Wednesday July 8, two days ahead of the broader launch. The architecture is full-duplex, meaning the model can listen and speak at the same time, register small acknowledgements like “mhmm” and “yeah,” and stay quiet when the user pauses. In head-to-head evaluations, GPT-Live-1 and GPT-Live-1 mini were strongly preferred over the older Advanced Voice Mode, with GPT-Live-1 also outperforming it on expert-level scientific reasoning and on agentic web search. GPT-Live routes deeper questions to GPT-5.5 in the background today.

According to OpenAI’s GPT-Live launch announcement, the system is rolling out globally on iOS, Android, and chatgpt.com from July 9. Free users get GPT-Live-1 mini; paid users get the full GPT-Live-1. Video and screen sharing are not supported at launch; subscribers continue to use Advanced Voice Mode for those features. GPT-Live carries its own safety stack layered on top of the model’s defences, including safeguards that can pause a conversation in higher-risk cases such as self-harm. The system was trained to refuse voice impersonation, and it stays inside a library of nine predefined ChatGPT voices.

The same release notes retire the Atlas browser on August 9, 2026, fold the App Directory into a new Plugin Directory, and stop creating new ChatGPT group chats on web, iOS, and Android from July 9. Existing group chats will become read-only at a later date. GPT-5.5 Instant Mini replaces GPT-5.3 Instant Mini as the fallback model users reach after hitting their GPT-5.5 Instant or Auto rate limits. None of those changes match the scale of GPT-5.6 and ChatGPT Work, and OpenAI bundled all of them into the same release notes page.

Frequently Asked Questions

What is ChatGPT Work and what does it do?

ChatGPT Work is the agentic mode OpenAI launched inside ChatGPT on July 9, 2026. According to OpenAI’s release notes, it handles connected apps and files into finished deliverables and runs Scheduled Tasks that fire once, on a schedule, or in response to a trigger. The agent also surfaces clarifying questions and approval prompts at consequential moments before it acts.

When will ChatGPT Work reach my account?

OpenAI’s release notes say ChatGPT Work rolls out first to Pro, Pro Lite, Enterprise, and Edu plans on web and mobile, with Plus and Business plans following over the coming days. Enterprise and Edu workspaces start with a two-week preview in which Work is off by default and admins can opt out before it auto-enables. Free and Go plans are not in the initial wave.

Why was the GPT-5.6 public release delayed?

On June 26, 2026, OpenAI said the U.S. government had asked it to start GPT-5.6 as a limited preview for a small group of trusted partners whose participation was shared with the government. The pause followed the Trump administration’s June 2026 executive order on frontier AI, which set up a voluntary pre-release review process for new models. Anthropic’s Claude Fable 5 and Mythos 5 had been disabled globally a few weeks earlier under a separate Commerce Department export control directive that was lifted in late June.

What’s the difference between Sol, Terra, and Luna?

Sol is OpenAI’s flagship for coding, science, cybersecurity, and complex reasoning, and it powers the Medium, High, and Extra High reasoning levels inside ChatGPT plus most of ChatGPT Work. Terra is the balanced everyday option, priced at $2.50 input and $15 output per million tokens. Luna is the fastest and lowest-cost model in the family, priced at $1 input and $6 output per million tokens, and it is available through Work, Codex, and the API rather than the ChatGPT picker.

How is GPT-5.6 different from GPT-5.5?

OpenAI calls GPT-5.6 Sol its strongest model yet, with state-of-the-art results on Terminal-Bench 2.1 for command-line workflows, gains on GeneBench v1 for biology, and cyber capabilities competitive with Anthropic’s Mythos Preview on ExploitBench. GPT-5.5 Instant remains the default model for fast ChatGPT responses. The new family introduces three durable tiers, Sol, Terra, and Luna, where the number marks the generation and the name marks the capability tier.