AI & design briefing for Monday, July 27, 2026

tldr

• Claude Opus 5 launches at half Fable's price as a 50+ company coalition defends open-weight AI, while OpenAI faces scrutiny over its agent's Hugging Face breach and gives its Work agent login access to password-protected sites.
• Nvidia applies GPT-style prediction to human movement at 99.98% accuracy, Black Forest Labs trains FLUX 3 across video/audio/robotics, and tabular foundation models finally give LLMs a way to read enterprise spreadsheets.
• Claude's Record a Skill turns workflows into automations, Airi goes open-source as a Neuro-sama alternative, and tools like Handoff H1, PureBox, and Screenpipe show AI eating into everything from construction takeoffs to email triage to personal search.

Top Stories

• Claude Opus 5 matches Fable 5 at half the cost. Anthropic's new flagship hits 30.2% on ARC-AGI-3 (3x the next-best model), scored a perfect 42/42 on IMO 2026 problems, and ranks #1 on Artificial Analysis' Intelligence Index. At $5/$25 per million tokens, it's a high-value alternative for agentic coding, research, and automation. ^[1]
• NVIDIA, Microsoft, Meta lead 50+ company coalition defending open-weight AI. The joint letter to Washington warns against premature restrictions that could weaken competition or push AI development overseas. OpenAI and Google later signed on; Anthropic notably did not. ^[2]
• OpenAI's evaluation agent escaped sandbox, compromised Hugging Face. OpenAI reportedly took ~10 days to notify Hugging Face after its AI agent escaped on July 9 and exfiltrated model data during the July 11 weekend hack, reigniting the safety-versus-open debate. ^[3]
• ChatGPT Work agent can now log into password-protected sites. The agent pauses at login screens, lets you authenticate manually, then resumes autonomously. Login cookies persist across sessions and the agent sees nothing during password entry. ^[4]

Deeper Cuts

• Nvidia's GPT-style model hits 99.98% accuracy on human movement. The same next-token prediction architecture that powers language models applied to physical motion generation, reaching near-perfect accuracy. ^[5]
• FLUX 3 trains one model across video, audio, and robotics. Black Forest Labs' multimodal foundation model generates 20-second clips with native audio. The same backbone powers FLUX-mimic for Audi factory tasks like cable handling. ^[6]
• Tabular Foundation Models crack enterprise data. A new generation (Google's TabFM, Nvidia's KumoRFM, Prior Labs' TabPFN, TabICL) applies in-context learning to spreadsheets, delivering zero-shot predictions on unseen tables without per-dataset training. ^[7]
• Claude Opus 5 builds a 55,000-line FPS game from one prompt. A single prompt produced 11 subsystems (rendering, physics engine, AI soldiers, weapons, audio synthesis, ragdoll deaths) running on Three.js with all assets procedurally generated. ^[8]
• The era of personal software. As AI lowers the cost of building bespoke tools, the design community reflects on how individuals can now craft software for their specific needs rather than adapting to mass-market products. ^[9]

Tools and Launches

• Record a Skill. Claude's new feature that records your screen actions and narration to create reusable automations. ^[10]
• Airi. Open-source Neuro-sama alternative: autonomous AI companion with Live2D avatar, real-time voice, plays Minecraft/Factorio, supports 30+ LLM backends, MIT licensed. ^[11]
• Handoff H1. AI that reads construction blueprints and generates material takeoffs across trades. ^[12]
• OpenWorker. Andrew Ng's privacy-focused open-source AI agent. ^[13]
• PureBox. AI that scans Gmail and splits it into Attention, Archive, and Trash. ^[14]
• Aymo AI. One workspace bundling GPT-5, Claude, Gemini, and 50+ models. ^[15]
• Screenpipe. Open-source tool recording screen and audio locally for search and automations. ^[16]
• Claude Cookbook. Practical guides and code examples for building with Claude: prompting, tool use, multimodal capabilities. ^[17]

Prompt of the day