OpenAI ships ChatGPT Work and GPT-5.6, taking direct aim at Claude Cowork

The July 9 launch pairs a long-running enterprise agent with a three-tier model family — Sol, Terra, and Luna — and names GPT-5.6 the 'preferred model' for Microsoft 365 Copilot.

OpenAI on July 9 launched ChatGPT Work alongside a three-tier GPT-5.6 family named Sol, Terra, and Luna, a coordinated release aimed squarely at Anthropic’s Claude Cowork and at the enterprise agent market Anthropic has been quietly compounding inside for the last year.

ChatGPT Work is the headline product. Per Bloomberg, it pulls context from apps, files, and workflows and returns finished documents, spreadsheets, presentations, reports, and even hosted websites, running across desktop, web, and mobile. Reuters reports the rollout began July 9 for Pro, Enterprise, and Edu tiers, with Plus and Business following in the days after. A new ChatGPT desktop application and a hosted-websites feature shipped the same day, which is less a coincidence than a scaffolding play: the agent needs surfaces to act on, and OpenAI is building them itself.

The model family underneath is where the strategic sequencing gets legible. Sol, the flagship, sits at $5 per million input tokens and $30 per million output tokens, and adds two reasoning modes (Max for deep single-threaded work, Ultra for coordinating subagents on long tasks). Terra, at $2.50 and $15, matches GPT-5.5 quality at roughly half the cost. Luna runs $1 and $6 for high-volume, low-latency work. GPT-5.4 retires July 23.

OpenAI says Sol “sets a new state of the art at 80, 2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about one-third less” on the Artificial Analysis Coding Agent Index. Sam Altman went further, telling CNBC Sol is “54% more token efficient on agentic coding.”

Then there’s the Microsoft question. Bloomberg previously reported Microsoft was swapping some OpenAI software in Word and Excel for its own MAI models. OpenAI’s counter-framing calls GPT-5.6 the “preferred model” for Microsoft 365 Copilot, though TechCrunch notes the term’s meaning is unclear and doesn’t necessarily contradict the earlier reporting. That’s narrative management at industrial scale.

The launch was itself delayed by a Trump administration national security review, with roughly 20 government-vetted partners getting early access under voluntary pre-release review. OpenAI’s own safety report rates all three tiers “High” risk for cyber and CBRN misuse, a disclosure that reads less like humility than like a permission structure being carefully documented in advance.

OpenAI ships ChatGPT Work and GPT-5.6, taking direct aim at Claude Cowork

Sources