a daily news desk
Tools

OpenAI ships GPT-5.6 and ChatGPT Work, its answer to Claude Cowork

A three-tier model family — Sol, Terra, Luna — and a long-running desktop agent land after a government-requested delay.

By Iris Holloway · Tools & frameworks ·

OpenAI released the GPT-5.6 model family and ChatGPT Work on Thursday, six months after Anthropic seeded the “agent that works for hours” category with Claude Cowork in January. Bloomberg and Reuters report the launch was held back at the request of an unnamed U.S. government agency on national-security grounds, a delay that reads less as friction and more as the new baseline for how frontier releases now clear the runway.

The model family splits three ways. Sol is the flagship, priced at $5 per million input tokens and $30 output, with a Sol Pro tier for complex work. Terra sits in the middle at $2.50/$15 for everyday tasks. Luna, the high-volume workhorse, comes in at $1/$6. TechCrunch confirms the ladder; OpenAI’s own post frames it as the deliberate segmentation the API market has been waiting for.

The benchmarks are pointed directly at Anthropic. Sol posts 92.2% on BrowseComp and 62.6% on OSWorld 2.0, and OpenAI claims it uses 85% fewer output tokens than Opus 4.8 on OSWorld. Sam Altman told CNBC the model is 54% more token-efficient than its predecessor on agentic coding. TechCrunch adds that OpenAI is citing the Artificial Analysis Coding Agent Index to argue it’s overtaken Anthropic there too. Efficiency, not raw capability, is where the argument is being fought now.

ChatGPT Work is the agentic layer on top. Reuters frames it as OpenAI’s direct answer to Claude Cowork. It pulls context from Slack, Notion, Microsoft 365, Google Drive, and local files, produces documents, spreadsheets, presentations, and web apps through a built-in browser, and can publish hosted sites directly. Rollout is staged: Pro, Enterprise, and Edu get web and mobile first, with Plus and Business to follow. Mac and Windows desktop apps ship to all tiers.

OpenAI says roughly 700,000 A100e GPU hours of automated red-teaming preceded release. That number, disclosed on the eve of a widely reported IPO, is doing narrative work of its own.

Sources