Morning Digest, July 9, 2026

13 newsletters, 7 overlapping stories

Top Stories

SpaceXAI and Cursor release Grok 4.5, an “Opus-class” model

(4 newsletters)

Grok 4.5 is the first model SpaceXAI and Cursor trained jointly after the $60B acquisition, and it lands with benchmarks in the range of Claude Opus 4.8 and GPT-5.5 while running around 80 tokens per second. The headline is efficiency and price: SpaceXAI claims roughly 4x the token efficiency of Opus 4.8, with pricing of $2 per million input tokens and $6 per million output versus $5 and $25 for Opus. Musk framed it as “Opus-class but faster, more token-efficient and lower cost,” with a larger model teased for next month, and usage is temporarily free inside Cursor and Grok Build.

Claude Cowork comes to web and mobile

(4 newsletters)

Anthropic extended Claude Cowork beyond the desktop so sessions and files are accessible across web and mobile, letting a long-running task start at your desk and be checked from your phone. The update also adds scheduled tasks that run even when your computer is closed. It is rolling out in beta over the coming weeks, starting with Max subscribers.

OpenAI upgrades ChatGPT voice with GPT-Live

(3 newsletters)

GPT-Live is a full-duplex voice architecture that listens, reasons, and speaks at the same time, removing the turn-based awkwardness of the old Advanced Voice Mode and enabling things like real-time translation. When a question needs more thinking, Live can hand the harder work to a stronger model in the background and keep the conversation going, sustaining chats for at least an hour. In OpenAI’s tests users preferred GPT-Live over the previous mode in roughly three of four head-to-head comparisons.

Google expands managed agents in the Gemini API

(3 newsletters)

Google added background execution, remote MCP server connections, custom function calling, and mid-session credential refresh to Managed Agents in the Gemini API. Developers still hit a single endpoint while Gemini handles reasoning, code execution, package installation, and file and web access inside an isolated cloud sandbox. The effect is to turn managed agents into asynchronous workers that run in real development environments without blocking the calling application.

Meta debuts its first image and video models

(2 newsletters)

Meta’s Superintelligence Labs launched Muse Image inside Meta AI, Instagram Stories, and WhatsApp, with a Muse Video preview to follow. The image model ranked #2 on Arena’s text-to-image leaderboard behind only GPT-Image-2, and it leans on Meta’s social graph so users can upload photos, mention friends, or pull from trending content. It is free for general use, with more advanced options in Meta’s paid tiers.

Harness engineering emerges as the frame for self-improving AI

(2 newsletters)

Lilian Weng’s long piece argues that the “harness,” the system around a base model that orchestrates thinking, tool calls, context, memory, and evaluation, is where recursive self-improvement actually happens. Design patterns like workflow automation, persistent memory, and sub-agents are what let models improve their own training and deployment loops. The open problems remain evaluation and managing context and memory well.

Also Worth Knowing

OpenAI will launch GPT-5.6 Sol, Terra, and Luna this Thursday. Sol is the flagship, Terra a balanced everyday model at roughly 2x cheaper than GPT-5.5, and Luna the fast low-cost option. (2 newsletters)
Microsoft is swapping OpenAI and Anthropic models for its own in Excel and Outlook. The move signals progress on cheaper in-house models as discounted token deals near expiry.
DeepSeek plans to make its own inference chips. The goal is to cut reliance on both Huawei and Nvidia amid US export controls.
Meta is building its first major Canadian data center. The 1-gigawatt Alberta facility will cost around $9B and take two to three years.
Blue Origin valued at $130B in its first outside funding round. Bezos is contributing $2B as the company targets a New Glenn return to flight by year end.
Cloudflare’s Monetization Gateway proposes a paywall for the agent internet. With over half of web traffic now non-human and AI agent requests up 1,700% year over year, the platform charges agents each time they tap a site’s resources.
GLM 5.2 signals a coming AI margin collapse. Open-weights models close to Opus quality now cost under 20% of Opus pricing through compatible endpoints, pressuring frontier inference margins.
TypeScript 7.0 ships a native compiler. Native code speed, shared-memory multithreading, and other work yield 8x to 12x speedups on full builds.
ByteDance targets design work with Seedream 5.0 Pro. The image model adds layer-separated precision editing and native support across 10-plus languages.
DoorDash open-sourced DashBench for AI code review. Replaying historical PRs, a Kimi K2.6 plus Claude Fable 5 combo beat DoorDash’s own production setup, with the takeaway that no single model dominates. (2 newsletters)
HubSpot details the infrastructure behind 20B-plus vectors. Moving from Helm to Kubernetes operators cut cluster spin-up from hours to minutes across 140-plus clusters.
A Bun developer rewrote Bun in Rust using Anthropic’s models. The agentic experiment produced a runtime that is faster, smaller, and lighter on memory, a job estimated at a year for a small team.

Quick Hits

Samsung out-earns Nvidia and Apple: an estimated $58B quarterly operating profit, up roughly 1,810% year over year, though the stock still fell 7%.
Fable 5 access extended: Anthropic is keeping Claude Fable 5 on paid plans through July 12 before it moves to usage credits, and is giving open-source contributors six months of Claude Max 20x.
Waymo expands: driverless rides are rolling out in San Diego, Las Vegas, Tampa, and Denver, starting with employees. Link
Prime Intellect raised a $130M Series A after crossing $100M in annualized sales in year one, and MiniMax is reportedly targeting a 2.7T-parameter model for Q3.
RAM crunch: analysts expect memory prices to jump 40-50% in Q3 2026 alone.
USPS Forever stamps rise to 82 cents from 78 on Sunday, the sixth hike in five years. Link
The private racetrack boom: ultra-wealthy car owners are trading country clubs for members-only circuits, with initiation fees running from $60k to $450k. Link
iPhone and fertility: studies tie up to 52% of the 2007 to 2011 US birthrate decline to the smartphone’s arrival.
beehiiv’s zero-to-one: a deep dive on how relentless product velocity, engineers with taste, and hand onboarding took the newsletter platform from underdog to market leader.

Shower Thoughts

Humans are easy to navigate around in a crowd because you can instantly tell where someone is going: faces point forward, backs mean they are moving away. If we moved sideways like crabs with no clear front, crowded places would probably turn into total gridlock. Source