OpenAI releases GPT-5.6, merging ChatGPT and Codex into a unified desktop application GPT-5.6 adopts a brand new three-layer capability system: Sol flagship model: Input $5 per million tokens, output $30, at the same price as the previous generation, supporting complex reasoning and long-term autonomous task execution; Terra offers high cost-effectiveness: Input costs $2.5 per million tokens, output costs $15, performance is close to GPT-5.5, but cost is halved, suitable for daily professional work; Luna's fastest and cheapest lightweight model: $1 per million tokens for input and $6 for output. Among them, Sol approaches or surpasses Claude Fable 5 in multiple agentic tasks and coding benchmarks. New inference mode: Max: Let the model invest more time in deep thinking, validation, and iteration. Ultra: Unique to Sol, it can call multiple sub agents in parallel to handle complex workflows. GPT-5.6 also adds Programmable Tool Calling, which supports models to run JavaScript in isolated environments to complete tasks such as computation and data processing; Prompt Cache Breakpoints， Further reduce the cost of enterprise deployment. The original ChatGPT application and Codex application have been merged into a unified desktop application, which includes three modes: Chat, Work, and Codex, sharing models, context, and task history. ChatGPT Work can connect with tools such as Google Drive, Slack, and email to independently disassemble tasks, execute work across applications, and generate documents PPT、 Final results such as tables, websites, etc. Sites can publish visualizations, dashboards, or widgets generated during ChatGPT conversations as websites with just one click; The browser control capability has also been upgraded synchronously, adding login mode website operation, multi tab page and file download capabilities, further enhancing AI's Computer Use ability.