GPT-5.6 built a voxel Manhattan after coding alone for nearly a week

Signals Inbox·July 10, 2026·AI DevTools

A creator says GPT-5.6 spent almost a week autonomously building a voxel version of Manhattan from one initial prompt. The signal? An AI coding agent apparently kept planning, coding, testing, and fixing its own work for days without being constantly rescued by a human.

Q1What actually happened?

Matt Shumer posted a demo claiming GPT-5.6 Sol created a detailed voxel version of Manhattan after running autonomously for almost a week. He describes it as a one-shot result, meaning the agent received the main objective and then handled most of the execution itself.

Q2What does one-shot really mean here?

It probably does not mean one model response. A week-long project would involve thousands of individual actions: writing files, running the program, checking errors, changing geometry, and trying again. One-shot is better understood as one initial assignment with no major human steering afterward. That is still impressive, but it is very different from producing Manhattan in a single magical answer.

Q3Why does the week-long runtime matter?

Because AI coding agents are usually better at short jobs than long projects. They can fix a bug or build a small app, but they often lose track of decisions, repeat mistakes, or break existing features after working for too long. An agent that remains productive for nearly a week is showing stamina, not just coding speed. That is one of the biggest missing pieces in autonomous software development.

Q4How does this compare with actual coding benchmarks?

The gap is still large. SaaSBench found that more than 95% of agent failures happened before reaching deep business logic, often during setup and integration. RoadmapBench tested realistic software upgrades involving a median of 3,700 changed lines across 51 files. Its strongest model completed only 39.1% of tasks. Long projects remain far from solved.

Q5So what may GPT-5.6 be improving?

OpenAI describes GPT-5.6 Sol as a model built for stronger coding, planning, and long-horizon agentic work. This demo appears to test exactly that promise. The important improvement may not be prettier code. It may be the ability to preserve a plan, inspect previous work, recover from errors, and keep moving toward the same goal for days.

Q6What could this change?

Vibe coding could move from helping people make quick prototypes to handling entire creative projects. Instead of prompting every few minutes, a user could describe a world, leave the agent working, and return days later to something explorable. That would make persistence and compute budgets more important than typing speed. The valuable skill becomes choosing the right goal and judging the finished work.