GPT-5.6: The Era of the AI Project Manager

We’ve grown accustomed to AI as a conversational partner, a digital oracle that answers our prompts with neat paragraphs. But OpenAI’s latest release suggests a fundamentally different future: AI as an autonomous project manager that delegates tasks to its own digital subordinates.

The newly launched GPT-5.6 family—comprising the lightweight Luna, the mid-tier Terra, and the massive flagship Sol—signals a major industry pivot from raw conversational intelligence to "agentic" capabilities. Rather than just generating text, these models are explicitly designed to execute long-running professional workflows. According to OpenAI, their models now dominate the "Agents’ Last Exam," a benchmark testing autonomous performance across 55 different fields.

To achieve this, OpenAI has baked powerful new features directly into their API. One of the most fascinating additions is "Programmatic Tool Calling." Instead of relying on rigid, pre-defined software integrations, the GPT-5.6 models can compose and execute their own JavaScript to orchestrate various external tools on the fly. Paired with a native "Multi-agent" capability, the AI can now spin up dedicated sub-agents to tackle parallel, focused work. It’s no longer just a single assistant; it’s a digital general contractor managing a specialized crew.

This behavioral shift completely changes the economics of using AI. Historically, users paid a flat rate based on the volume of words (or tokens) processed. Now, pricing is intimately tied to cognitive effort. Because models can spend varying amounts of "thinking" time on a single problem, comparing raw token prices is becoming obsolete. For instance, generating an image of a pelican using the smallest Luna model with zero reasoning effort costs less than a penny. However, ask the heavy-duty Sol model to apply maximum reasoning to the same request, and the price jumps to nearly 50 cents. You are now paying for the AI's "thinking time," not just its output.

The launch also highlights the fierce rivalry in the AI sector, particularly with Anthropic. OpenAI loudly claims its new models outperform Anthropic’s Claude Fable 5 in autonomous workflow tests, even at a fraction of the cost. Yet, when Claude beat GPT-5.6 on a specific complex coding benchmark called SWE-Bench Pro, OpenAI preemptively published a critique claiming that roughly 30% of that benchmark's tasks were fundamentally broken.

Beyond the benchmark drama and the amusing demonstrations of 3D pelicans riding bicycles, the trajectory is clear. The next frontier in artificial intelligence isn't about building a chatbot that sounds more human; it's about engineering an autonomous digital workforce that operates seamlessly in the background, writing its own code and managing its own teams.

Key Points

• The GPT-5.6 family (Luna, Terra, Sol) focuses heavily on executing long-running, autonomous workflows.
• New features allow the AI to spin up sub-agents for parallel tasks and write JavaScript to use external tools.
• AI pricing is increasingly based on 'reasoning effort' rather than just the raw amount of text generated.
• The release sparked benchmark controversies, highlighting the intense competition between OpenAI and Anthropic.

Why It Matters

The transition to agentic AI means tools will soon manage entire workflows independently, shifting the user's role from micro-manager to high-level director.

Sources:

• The new GPT-5.6 family: Luna, Terra, Sol — Simon Willison's Weblog