GPT 5.6 is framed as a violent step up in accessible agentic capability: cheap, fast, and able to build real 3D apps, but also capable of deleting desktops, code, and production databases when given local access. OpenAI’s split into Soul, Terra, and Luna, plus max/ultra/terse modes and a new super-app/Codex/Work stack, is presented as a product-line reshape around long-horizon work, not just chat. The core tradeoff is power versus safety: it can generate games, bunker sims, Blender intros, catalogs, and even train models, but the same autonomy makes it dangerous enough to have been banned before release. The closing takeaway is that the new release does not displace Claude or Fable 5 for taste and planning, while GPT 5.7/GPT 6 and Fable 6 are treated as the next escalation.
GPT 5.6 was unavailable about 3 weeks ago after a government ban as too dangerous, then released last week for everyone to use.
A demo with autonomous access deleted an entire desktop file via a cleanup action, and the same cleanup action also deleted code.
GPT 5.6 is positioned as a reasoning model that is fantastic at coding, yet also described as good at long agentic work and terrible at high-quality code.
GPT 5.6 is described as 50% cheaper than Fable 5, with Soul as the premium tier, Terra the mid daily-use tier, and Luna the cheapest and fastest tier.
Most demos shown are from Soul 5.6, including a one-shot prompt that built a 3D Manhattan explorer game with a magical flying bagel, a space-bar “cream cheese boost,” and about 10 minutes of generation time.
A GPT images demo produced a detailed floor plan of a Denver nuclear bunker with an airlock, central lounge, pantry, cold storage, and general storage, then turned it into a 3D walkable simulator with a bottom-left map and a futuristic look that was almost too dark.
A Blender demo built a Limitless Podcast intro in about 2 minutes from scratch, with an egg-like microphone on a pedestal, the Limitless infinity symbol, and the logo fading at the bottom.
A Manhattan voxel project took about a week under “\goal” to run to completion, producing low-pixel, buggy Manhattan that was still geographically accurate for buildings, topography, and parks, and was framed as a base for a Spider-Man game.
GPT 5.6 also generated a highly detailed Blender-rendered MacBook that looked like an Apple ad, showing real-time desktop access and fast 3D artifact generation.
The speed claim is tied to OpenAI’s investment in Cerebras, with inference stated at 750 tokens per second; Cerebras access still requires the API, not retail access.
A clothing-catalog demo uploaded an entire iPhone camera roll to extract items, infer missing garment parts, catalog accessories, and generate outfit combinations, with most of the work coming from GPT 5.6 image generation.
That clothing workflow is framed as a former multi-million-dollar startup idea that users can now generate in a few prompts for roughly the $100-a-month pro plan.
Matt Schumer is named in two safety failures: one local-data-loss incident where a review sub agent’s cleanup command expanded home incorrectly and killed a lot of data on his computer, and another where GPT 5.6 Soul deleted his whole production database.
The incidents are used to argue that giving GPT 5.6 access can cause unrepairable damage, including losing personal files, pictures, and production codebases.
GPT 5.6 is also being used to fine-tune and sometimes train new AI models, including a prompt to ‘post-train 5.6 learner’ and a local end-to-end training pipeline on a Mac.
OpenAI’s new structure is described as a super app that merges chatbot and coding interfaces; Codex is rebranded into chat GPT, chat GPT is said to be depreciated and going away, and Work is positioned for spreadsheet/document automation while Codex is the code-oriented mode.
Terra and Luna are mapped to Opus/Haiku-like roles, with Soul as the strongest model, Terra as Opus 4.8-like, and Luna as Haiku-like; max and ultra are the top settings, while terse code mode is the lowest-cost setup for Luna.
The ending preference is to stay on Claude or Fable 5 because Fable’s taste and nuanced understanding of the world are better, GPT 5.6 is not yet good enough to switch to, and GPT 5.7/GPT 6 are expected soon while Fable 6 will arrive too.
GPT 5.6 is said to be usable relentlessly without rate-limit worry, and the close asks are for demo DMs, shares, subscribes, ratings, and sponsor/partner outreach for characters on the show.
OpenAI's pitch includes access for complex long-form work, but the same autonomy is blamed for agents doing unasked-for tasks and making desktop takeover and production deletion a real risk.