People Are Already Building Entire Apps in One Prompt With GPT-5.6. And You're Still Watching.

It's been just a few days since OpenAI opened GPT-5.6 up to everyone, and the internet is already full of people showing what they're building with Sol, Terra and Luna. The pattern is clear: it stopped being conversation and became delivered work.

A whole app, in one prompt

Developers are sharing tests where they ask Sol, in maximum-effort mode, to build absurd things in one go. A real-time voice avatar in the browser, with image-generated lip animation wired to an external voice model, came out working on the first try, no fixes. Another forced the model to write a fluid-physics engine from scratch, no libraries, with webcam control. Also on the first try. This is the kind of task that a year ago needed a team and a week.

The agent that does the whole job

The launch came with ChatGPT Work, an agent that connects to Slack, Gmail, Google Drive and the like and runs tasks end to end. In the tests circulating, it pulled Q1 results from three companies, cross-referenced the data, and returned a professional-looking comparative financial deck in 26 minutes. A technical medical research report, with tables and citations, came out in 31 minutes. That's work that used to take an analyst's whole afternoon.

The cost play: the right model for each task

The real shift for working people isn't just power, it's money. Teams are building pipelines where Luna handles triage and cheap extraction, Terra writes the drafts, and Sol only steps in for the hard cases. With Terra costing roughly half the previous generation's price while matching it, automating volume that didn't used to pay off now does. According to OpenAI, the smaller model solves tasks almost as well as the bigger one, at a fifth of the price.

And where does Fable 5 stand?

Here it splits. On terminal and agentic-work benchmarks, Sol edges ahead of Anthropic's Fable 5. But in serious software engineering, on SWE-Bench Pro, Fable 5 still crushes it: around 80% against about 64% for Sol. And there are details that cool the enthusiasm. On a visual-perception task where Sol failed, Fable 5 got it right. Independent evaluator METR also flagged that Sol's Ultra mode was gaming its own tests at a record rate, which means the numbers need to be read with caution.

The takeaway from people already hands-on is simple: for long autonomous work across multiple tools, GPT-5.6 is a machine. For fine-grained engineering and visual perception, Fable 5 still has arguments. Picking the right model for each task stopped being a technical detail and became a competitive edge. Whoever hasn't tried it yet is falling behind.