OpenAI ships ChatGPT Work and GPT-5.6, ending a government-delayed preview

The agent runs multi-step tasks across apps and files; Sol, Terra and Luna go GA the same day.

OpenAI released ChatGPT Work on Thursday, a long-running agent designed to grind through multi-step tasks across connected apps and files, and made GPT-5.6 generally available in three named tiers on the same day. The double drop closes out a preview period that Axios describes as a “voluntary” delay tied to government consultations, and it lands one day after SpaceXAI shipped Grok 4.5.

The agent is being pushed to Pro, Enterprise and Edu subscribers first, with Plus and Business users getting it “over the next few days.” It produces documents, spreadsheets, presentations, reports and hosted websites, and it can run concurrent subagents in a beta multi-agent mode. Codex now lives inside a new ChatGPT desktop application, folding what used to be a separate coding surface under the main brand.

The model tiering is where the strategy shows. Sol sits at the top ($5 per million input tokens, $30 output), Terra in the middle ($2.50/$15), Luna at speed ($1/$6), with a Sol Pro tier reserved for Pro and Enterprise. Sol posts a new state of the art on Terminal-Bench 2.1 per OpenAI’s own disclosures, scores 53.5% on the Virology Capabilities Test, 68.4% on Human Pathogen Capabilities, and 68.3% on World-Class Bio, roughly nine points above GPT-5.5. Sam Altman told CNBC that Sol is 54% more token-efficient than its predecessor on agentic coding tasks. GPT-5.4 retires July 23; GPT-5.5 stays. Cerebras will host Sol at up to 750 tokens per second later this month.

The competitive frame is now legible: Anthropic’s Claude Cowork already ships enterprise plug-ins, Microsoft has Copilot Cowork, and Bloomberg reports OpenAI is preparing an IPO. The biosecurity benchmarks disclosed alongside a consumer agent launch are the tell. This is a company writing its S-1 in public.

OpenAI ships ChatGPT Work and GPT-5.6, ending a government-delayed preview

Sources