AI & design briefing for Monday, July 13, 2026

20 stories merged from 4 newsletters

tldr

  • Apple sued OpenAI over alleged hardware trade-secret theft, GPT-5.6 Sol Ultra proved a 50-year-old math conjecture with 64 parallel subagents, and OpenAI's safety head departed.
  • Model routing emerges as critical middleware, DeepSeek and Meta both push custom AI chips, and Thinking Machines Lab argues for human-centered AI over full autonomy.
  • Cursor ships side chats, Grok 4.5 launches at $2/$6 per million tokens, and ByteDance's Seedream 5.0 Pro undercuts GPT Image 2 by ~9x on price.

Top Stories

Deeper Cuts

  • Model routers as the missing middleware. AlphaSignal's deep dive covers heuristic routers, learned classifiers, cascades, and ensembles. OpenRouter's Fusion beat standalone GPT-5.5 on the DRACO benchmark at roughly half the token cost; open-source options like NadirClaw cut API bills by up to 70%. [6]
  • DeepSeek reportedly developing inference-focused AI chip. Reuters reports DeepSeek has spent ~a year designing an inference chip and is talking to foundry partners. With ~70% of AI compute demand being inference, a lab designing both models and chips could squeeze Nvidia's China business. [7]
  • Meta's Iris AI chip enters production in September. Designed with Broadcom, manufactured by TSMC. Meta plans 7GW of compute capacity in 2026, doubling to ~14GW in 2027 on $125-145B spend. Broadcom holds a $73B AI chip backlog across Google, OpenAI, Meta, and ByteDance. [8]
  • Microsoft routes some Copilot prompts to internal models. GPT-5.6 became the preferred model in Microsoft 365 Copilot, but Microsoft is routing some Excel/Outlook prompts to internal models to cut inference costs. [9]
  • Thinking Machines Lab argues for human-centered AI. The lab's essay rejects the autonomy-timescale progress measure, arguing users should train and own their models, doubling as the case for its Tinker product. [10]
  • AI 2040 drafts humanity's "Plan A". The AI Futures Project's new five-scenario forecast recommends US and China pausing frontier training in 2029 with expanded chip tracking. Co-author Daniel Kokotajlo predicts "companies won't like it." [11]

Tools and Launches

  • Cursor v3.11. Side chats (/side, /btw) for parallel conversations, searchable agent history with Cmd+F, and cloud agent hooks for self-correcting loops. [12]
  • Grok 4.5. xAI and Cursor co-developed model at $2/$6 per million tokens, trained on trillions of tokens of Cursor data; claims to beat Claude Opus 4.8 on coding benchmarks. [13]
  • Cognition SWE-1.7. Coding model built on Kimi K2.7, runs at 1,000 tokens/sec via Cerebras, free in Devin's standard tier. [14]
  • Runway Dev. Single API serving Runway's frontier video models alongside third-party image and audio generators. [15]
  • Mistral Robostral Navigate. Mistral's first robotics model for plain-language robot navigation across factories and warehouses. [16]
  • GitHub Spec Kit. Spec-first workflow forcing requirements, clarification, and planning before coding agents implement. [17]
  • ByteDance Seedream 5.0 Pro. Image model with layer separation, infographic generation, and 10+ language text rendering at ~$0.03/image vs ~$0.28 for GPT Image 2. [18]
  • Google AI Studio custom domains. Deployed apps can claim a permanent subdomain under ai.studio instead of a random URL. [19]
  • a11y.quest. 128 practice questions for WCAG 2.2, ARIA, semantic HTML, keyboard, and contrast with plain-English answers. [20]

Sources: The Neuron, The Rundown AI, AlphaSignal, The Strange Review