Tuesday, July 14, 2026

OpenAI Opens The Floodgates

OpenAI Opens The Floodgates

Today’s Overview

Good morning, OpenAI is loosening GPT-5.6 Sol limits after a surge in demand, while Anthropic is pushing Claude Cowork onto phones and the web for work that keeps moving without your laptop. Meanwhile, Microsoft’s AI security agents are turning bug discovery into a Patch Tuesday pressure cooker. Let’s dive in.

Top Stories

OpenAI Temporarily Relaxes GPT-5.6 Sol Limits

OpenAI is temporarily relaxing GPT-5.6 Sol usage limits after demand surged for its most powerful model. The change removes the five-hour usage restriction for Plus, Pro, and Business plans and resets current usage for users, giving people more room for coding and agentic work.

  • The change applies across Plus, Pro, and Business plans, rather than only the highest-paid tier.
  • Codex and ChatGPT usage are linked through shared limits covering local messages and cloud-based tasks.
  • OpenAI says GPT-5.6 Sol is being made more efficient so less usage is consumed for the same work.

Claude Cowork Comes To Mobile And Web

Anthropic is bringing Claude Cowork, its multi-step work agent, to mobile and web. The move makes Cowork more useful for asynchronous tasks across files, email, calendars, messaging apps, the web, and connected tools, with scheduled work continuing even when no device is online.

  • Beta access is rolling out over several weeks starting with Max users, with more plans to follow.
  • On desktop, Claude can also use local files and browser for the full Cowork experience.
  • Anthropic is extending doubled Cowork usage limits through August 5 to encourage larger delegated tasks.

Microsoft’s AI Bug Hunters Raise Patch Tuesday Stakes

Microsoft has built MDASH, a multi-model agentic scanning harness that coordinates more than 100 specialized AI agents to find and prove exploitable bugs. The system is already changing the security workflow by finding vulnerabilities faster, while Satya Nadella is also warning that enterprise AI can leak organizational know-how unless memory and learning loops stay inside customer-controlled boundaries.

  • Microsoft says MDASH helped find and fix 16 vulnerabilities ahead of a Patch Tuesday release.
  • The system scored 88.45% on CyberGym across 1,507 real-world vulnerabilities.
  • In retrospective testing, MDASH reached 96% CLFS recall and 100% recall on tcpip.sys cases.

Research & Analysis

LightMem-Ego Brings Everyday Memory To Wearables

LightMem-Ego is a lightweight streaming multimodal memory system designed for personal AI assistants on phones and wearable devices. It aligns egocentric visual and audio streams on a shared timeline, organizes them into current, short-term, and long-term memory, and routes each user query to the right memory level for grounded answers.

  • The system is aimed at everyday-life assistance rather than one-off image or audio understanding.
  • Its demo scenarios include object finding conversation recall, life summarization, routine discovery, and personalized help.
  • The paper says the project includes available code through the linked repository.

Northwestern Gives AI A Cerebellum-Like Upgrade

Northwestern describes research that adds a cerebellum-like component to AI systems. The work is presented as a technical research finding from the university, not a product launch.

  • The university frames the work around brain-inspired AI using the cerebellum as the reference point.
  • The story appears in Northwestern Engineering’s research news rather than a commercial product channel.
  • The source material supports treating it as early technical research with limited implementation detail provided.

ABot-N1 Targets General Visual-Language Navigation

ABot-N1 is presented as a step toward a general Visual Language Navigation foundation model. It separates cognition from control with a slow vision-language reasoner that produces a pixel goal and a fast action expert that turns textual cues and pixel guidance into continuous waypoints.

  • The model uses image-space anchor points as a shared interface across navigation tasks.
  • Its task coverage spans point-goal and object-goal navigation, POI-goal, instruction-following, and person-following.
  • The paper reports 95.4% and 92.9% SR in complex indoor and outdoor scenes.

Sakana AI Builds Smart Cellular Bricks

Sakana AI is extending its collective-intelligence research into physical hardware with Smart Cellular Bricks. The bricks use local communication and neural networks to classify and reconstruct 3D shapes without centralized control.

  • The project moves collective intelligence into physical modules instead of keeping it purely in simulation.
  • The system relies on local communication so the structure can operate without a central controller.
  • Experiments tested robustness against noise and module failures while maintaining shape-classification performance.

Trending AI Tools

  • Colibri Runs GLM-5.2, a 744B MoE model, on a 25GB RAM machine with no GPU by keeping the core in memory and streaming the rest from disk.

  • Google AI Studio custom domains Lets deployed apps use unique ai.studio subdomains instead of random run.app URLs.

  • Prime Intellect verifiers v1 Introduces an interception architecture for verifier-oriented agent training across competing model APIs.

Quick Hits

  • Blender MCP bridges AI assistants with Blender so users can create, edit, light, and render 3D scenes through natural language.

  • Tom Blomfield joins Anthropic on a leave of absence from YC to work with Tom Brown on the compute team.

  • Claude Fable 5 access extended through July 19, alongside elevated Claude Code limits and OpenAI’s temporary GPT-5.6 Sol cap removal.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.