Key Points
- OpenAI's GPT-5.6 Soul launch went viral on the strength of multimodal demos—users recreated *Interstellar* in 3D and built Blender scenes in real time at 750 tokens per second, creating a visceral sense of machine autonomy that resonates beyond raw benchmarks.
- OpenAI consolidated its Codex and ChatGPT desktop apps into a single Electron-based application, eliminating native integrations and command-line shortcuts that power users relied on, forcing migration to ChatGPT Classic.
- OpenAI acknowledged the launch bungled clarity around compute settings and usage limits, reset usage caps twice on day one, and signaled uncertainty about whether Codex and ChatGPT should remain separate products rather than merge into one omnibox.
Summary
GPT-5.6 Soul Launches to Viral Demos and Desktop App Backlash
OpenAI released GPT-5.6 Soul to immediate viral attention driven by striking visual demos, but the launch triggered friction over the new desktop app architecture and some clarity issues around usage limits.
The headline demos showcase the model's speed and multimodal capability. A user recreated the entire film Interstellar as a 3D scene from prompt to rendered output, illustrating how quickly the model can move from text to three-dimensional models. In a separate demo, GPT-5.6 Soul built a three-dimensional cannon in Blender, adding primitives and textures in real time without video acceleration. According to one observer, the model achieved approximately 750 tokens per second in these conditions—benchmarking against Cerebrus, though the performance came from SOL Ultra running in fast mode.
The speed itself signals a tangible threshold: watching the model interact with desktop software at human-impossible rates creates a distinct impression of capability. One speaker notes the psychological weight of observing a machine use a computer at speeds that would be unfeasible for a person unless software interaction was their entire job.
These demos also exposed a creative use pattern. Users are combining blocky, low-fidelity mockups—either iPhone recordings of actors delivering lines or Blender placeholder scenes—with AI video models to generate higher-fidelity output. The reference video stays physically deterministic, which appears to help the model avoid hallucination drift on longer sequences.
The practical upside for design work proved notable. One user reported that GPT-5.6 Soul generated a web design logo without being asked, unprompted improving on previous attempts. The model now integrates image generation directly into design workflows, producing preview images for websites it builds. The underlying image model now has tool use and can scaffold HTML, blurring lines between the text, image, and coding capabilities.
One anecdotal story captured the scale of the model's autonomy: a user was banned from OpenAI's platform, pasted the ban notice into Codex, which analyzed it, identified that the user had requested an API key to their own server, wrote an appeal, submitted it, and received auto-approval from an AI system within ten minutes. The user had been "banned by AI, convicted by AI, defended by AI, and pardoned by AI in about ten minutes."
Desktop App Friction
The new ChatGPT desktop app consolidated Codex and ChatGPT Work into a single Electron-based application, replacing the previous native app. The shift created user friction: the old app offered native integrations and command-line shortcuts (command-plus for quick access on Mac). The new Codex app, built as web views, requires users who want the legacy experience to manually switch to ChatGPT Classic, a maintenance burden OpenAI is unlikely to support indefinitely.
The scope of actual daily desktop app usage is unclear. One speaker noted they primarily used the Codex app on desktop, opening a new tab for chat-based conversations, and used the ChatGPT mobile app to sync across devices. They rarely opened the desktop ChatGPT app directly. But for users with ingrained workflows built around native shortcuts and integrations, the migration imposes real friction.
OpenAI acknowledged the misstep. In a statement, the team said they "made it too easy to use the highest compute settings without making the impact on usage limits sufficiently clear" and "reorganized the desktop app in one bold move, making familiar things like chats and projects harder to find." They also noted their launch framing focused on work in a way that made it feel to Codex users like the product was being deprecated, which they denied. OpenAI reset usage limits twice on launch day to let users continue experimenting.
The confusion also reflects a deeper product question: whether ChatGPT Work and ChatGPT Codex should remain separate. One analogy that emerged is enterprise Google Search—it does not exist as a separate product from consumer Google. The same search box works everywhere. Under that model, a single omnibox that handles both chat and coding tasks would be the end state.
Every deal, every interview. 5 minutes.
TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.