# GPT-5.6 Sol runs a week-long voxel Manhattan build in Matt Shumer demo > Source: > Published: 2026-07-09 18:50:39+00:00 [Matt Shumer (@mattshumer_)](https://x.com/mattshumer_) said on July 9th that OpenAI's GPT-5.6 Sol generated a voxel-based Manhattan scene in a run that lasted almost a week and operated autonomously, a small public glimpse of the agentic workload OpenAI has been describing around its newest flagship model. [https://x.com/mattshumer_/status/2075268746315268138](https://x.com/mattshumer_/status/2075268746315268138) The claim came in [a two-post thread on X](https://x.com/mattshumer_/status/2075268746315268138), where Shumer posted a video of the voxel city and wrote that GPT-5.6 Sol "one-shotted" the build. The thread gives the useful part and the missing part in the same frame: Shumer says the model ran for nearly a week without human intervention, while the post includes no prompt, code repository, execution logs, environment details, token count, or cost figure. Shumer is not a random model tourist. He co-founded OthersideAI, the company behind HyperWrite, and has spent the past few years building and promoting consumer-facing AI agents. Forbes listed OthersideAI's co-founders as Shumer and Jason Kuperberg and described the startup as a generative AI company that embedded AI into websites for writing and simple browser tasks. The same Forbes profile said the company had received $5.8 million in funding and claimed nearly two million users as of its last update. In April, Shumer wrote that he and Kuperberg had stepped away from HyperWrite "a few months" earlier, with Josh Bickett taking over as CEO, and said the pair felt "the pull to start exploring again." That background matters because the demo is less a consumer product announcement than an operator's test of how far a frontier coding agent can run without being babysat. HyperWrite's original promise was browser automation and writing assistance. Shumer's July 9th post points at a harder version of the same thesis: a model that can stay on task across days, make thousands of small implementation decisions, and produce a coherent artifact at the end. OpenAI introduced GPT-5.6 Sol on June 26th as part of a three-model GPT-5.6 family: Sol as the flagship, Terra as the balanced lower-cost option, and Luna as the fastest, cheapest tier. In its [launch announcement](https://openai.com/index/previewing-gpt-5-6-sol/), OpenAI described Sol as its strongest model and said GPT-5.6 adds a new `max` reasoning effort plus an `ultra` mode that uses subagents for complex work. OpenAI also said the model family improves agentic capabilities in coding, biology, and cybersecurity. The access story is central to the demo. OpenAI's [Help Center article](https://help.openai.com/en/articles/20001325-a-preview-of-gpt-5-6-sol-terra-and-luna) says GPT-5.6 Sol, Terra, and Luna are available during preview only through the OpenAI API and Codex for a limited group of trusted partners and organizations. The same article says individual consumers cannot enroll, ChatGPT access is excluded during the preview, and OpenAI has not announced a general-availability date. That makes Shumer's post a distribution event as much as a technical one. The public can watch selected operators test the model on long-running tasks, while most developers cannot yet reproduce the test. The setup gives OpenAI controlled proof points from credible users, and it gives those users early social proof around agent workflows that competitors and customers are watching closely. The cost question is also unresolved. OpenAI lists GPT-5.6 Sol at $5.00 per 1 million input tokens and $30.00 per 1 million output tokens during preview, with Terra at $2.50 input and $15.00 output, and Luna at $1.00 input and $6.00 output. A near-week autonomous run can accumulate cost through long context, tool calls, retries, generated code, and validation loops. Shumer did not disclose what the Manhattan build cost. In a reply to [Ted Benson (@edwardbenson)](https://x.com/edwardbenson), Shumer wrote that this kind of task would be "one-shottable" within a year and added that it would cost "a stupid amount." The video should be read as a supplied demo rather than an audited benchmark. A model-generated voxel Manhattan is visually legible, but the thread does not show whether the model produced every asset from scratch, used a framework or template, recovered from failures, or required hidden setup. It also does not show the success rate across repeated attempts. For founders evaluating whether frontier agents are ready for production work, those details decide whether a week-long autonomous run is a capability they can buy or a stunt they can watch. Still, the demo lands at the exact pressure point in the current model race. OpenAI's own framing for GPT-5.6 Sol emphasizes long-horizon command-line workflows, planning, iteration, and tool coordination. A voxel Manhattan is not enterprise software, but it is a clean visual proxy for persistence: large structure, repeated patterns, fine detail, and enough room for errors to compound. If Shumer's description is accurate, the run shows Sol being used less like a chatbot and more like a background worker that can spend days producing an artifact. OpenAI has also been explicit that the same step up in long-horizon capability brings risk. Its launch post says GPT-5.6 Sol improves cybersecurity performance, uses layered safeguards, and remains below OpenAI's Cyber Critical threshold under the company's Preparedness Framework. The Help Center article says some preview requests can be blocked or delayed by additional safety checks, especially in biological and cybersecurity contexts. Creative coding demos are the benign side of the same product direction: models that can reason longer, call tools, coordinate subagents, and keep going. Shumer's Manhattan post does not settle whether GPT-5.6 Sol changes the economics of software creation. It does show why access to the model has become valuable before broad release. The people with preview access can test week-scale agent loops while the rest of the market waits for pricing, limits, reliability data, and the ordinary failure reports that come only after many developers try the same thing on their own jobs.