OpenAI on Thursday released the GPT-5.6 model family alongside ChatGPT Work, an agent designed to autonomously execute complex, multi-step workplace tasks across applications. The launch, which followed a weeks-long US government restriction on the model’s distribution, signals a structural shift in the artificial intelligence race: the frontier is no longer defined by conversational fluency but by the capacity to complete real-world projects—drafting documents, analysing spreadsheets, building web apps—over extended periods with limited human intervention. The flagship Sol model claims a 13.1-point lead over Anthropic’s Fable 5 on the Agents’ Last Exam benchmark and a narrow edge on the Artificial Analysis Coding Agent Index, while using less than half the tokens.

The mechanism behind this shift is a deliberate decoupling of capability and cost. GPT-5.6 is offered in three tiers—Sol, the mid-range Terra, and the budget Luna—priced from $1 to $30 per million output tokens, a structure that OpenAI’s CEO Sam Altman, speaking from the Sun Valley conference, linked directly to enterprise demands for efficiency and measurable return on investment. The new ChatGPT Work agent integrates the Codex developer toolset into the consumer interface, connects to over 1,400 plugins including Gmail, Slack, and Google Drive, and introduces a scheduling function for recurring tasks. A built-in desktop browser and a “Computer Use” feature allow the system to operate applications on a user’s behalf, moving files and entering text while the user is away.

Viewed from Washington, the rollout unfolded against a backdrop of regulatory friction. The Commerce Department had earlier forced Anthropic to suspend access to its Fable 5 and Mythos 5 models for 19 days, citing cybersecurity risks, and pressed OpenAI to limit GPT-5.6 to vetted partners. Microsoft President Brad Smith warned that the US government was regulating frontier AI “without transparent or complete rules,” a concern echoed by analysts in London who note that the episode has damaged international confidence in the stability of American AI supply chains. In Beijing, users accessing GPT-5.6 via VPNs praised its efficiency even though it remains more expensive than domestic alternatives from DeepSeek and Zhipu AI, which charge as little as $0.44 per million input tokens.

The competitive response was immediate. Anthropic had already expanded its Claude Cowork agent to web and mobile days earlier, while Meta released Muse Spark 1.1, a coding-focused model, and began charging for API access for the first time. The industry is consolidating around a new paradigm: AI systems that do not merely answer prompts but complete entire workflows. The next factual milestone is the phased rollout of ChatGPT Work to Plus and Business subscribers in the coming days, alongside the planned sunset of OpenAI’s standalone Atlas browser on 9 August, as the company streamlines its product suite ahead of an anticipated initial public offering.