6 min read · 14 sources

Anthropic has officially outgrown Git. The lab replaced the world’s default version control system with an in-house platform called Scion to prevent autonomous coding agents and engineers from grinding their repository to a halt. The migration took five months, accommodating a jump from 2,500 daily merges in May to more than 30,000 today.

Simultaneously, Google opened up access to Gemini 4 Argon with an aggressive pricing cut aimed directly at enterprise workloads, and the FBI arrested an OpenAI engineer for planting an encrypted bug inside the company’s workspace. Infrastructure bottlenecks, pricing wars, and corporate espionage all landed on the same morning.

Here is what you need to know.

Anthropic migrated its 100 GB monorepo from Git in five months, scaling merge throughput from 2,500 daily commits in May to over 30,000.

Anthropic ditches Git for Scion after hitting 30,000 daily merges

Source: anthropic.com ↗

Distributed version control systems break down when the commit log acts less like a human diary and more like an automated telemetry stream. Anthropic published architectural details on Scion, its homegrown source control system, which completely replaced Git for the company’s primary monorepo. The core issue: Git’s design presumes clients clone complete repository histories. With Anthropic’s main branch history sitting at roughly 100 GB, standard git fetch and checkout operations became an untenable operational tax.

The transition from Git took five months. Scion now handles over 30,000 merges per day - up more than tenfold from the 2,500 daily merges Anthropic logged in May 2026. This sudden acceleration is driven primarily by internal automated workflows and code-generation agents landing changes continuously. Alongside the storage engine, Anthropic built Antique (a custom code review interface), a monorepo search layer, an IAM system scoped to individual repository paths, and automated landing pipelines that process large volumes of accepted patches without head-of-line blocking. For infrastructure teams managing monorepos over 50 GB, Anthropic’s pivot reinforces a reality Meta and Google hit years ago: decentralized object graphs do not survive automated scale.

Google drops Gemini 4 Argon into broad availability with aggressive API cuts

Source: venturebeat.com ↗

Google’s frontier model, Gemini 4 Argon, is now in general availability across the paid Gemini API, Google AI Studio, Antigravity, and Gemini Enterprise. For individual developers, the model is exposed via Google AI Pro/Ultra tiers and AI Mode in Search.

To buy market share for coding, cybersecurity, and financial analysis workloads, Google slashed Argon’s API pricing to $2 per million input tokens and $10 per million output tokens for a limited window. That pricing undercuts comparable reasoning and multimodal frontier models from rivals. It is designed to entice engineering organizations to embed Argon into automated pipelines and deep analytical tasks where token volume usually makes production deployments cost-prohibitive.

FBI arrests OpenAI engineer over encrypted bug planted in office

Source: nytimes.com ↗

Federal law enforcement stepped into frontier AI security this week. The FBI arrested Abdulrahman Alfozan, an engineer on OpenAI’s product team, after physical surveillance equipment was discovered in an OpenAI office. Alfozan, a Saudi Arabian national who joined the company in July 2025, allegedly planted a concealed recording device designed specifically to evade detection sweeps.

According to investigators, the device recorded ambient room audio, utilized advanced encryption to protect stored data, and incorporated an exfiltration mechanism designed to transmit captured data without tripping internal network monitoring. The physical breach highlights that as frontier labs house increasingly sensitive model weights, pre-release evals, and system architectures, the threat model has fully shifted from standard network perimeter defense to physical espionage and insider threats.

SpaceX eyes 456 Starship flights per year in South Texas

Source: bloomberg.com ↗

Federal regulators are reviewing a filing from SpaceX proposing up to 456 launches and 912 landings per year at its Starbase facility in Boca Chica, Texas. Each orbital mission involves two landings: the Super Heavy booster returning to the launch mount chopsticks, followed by the Starship upper stage.

The regulatory push is the operational counterpart to Elon Musk’s goal of launching Starship on hourly cadences within the next decade. Alongside the Texas regulatory filings, SpaceX recently unveiled plans to invest up to $100 billion to construct a second massive launch complex in southern Louisiana. Operating hundreds of heavy-lift launches annually requires turnaround cycles closer to commercial aviation than traditional aerospace, demanding automated propellant loading, instant pad refurbishment, and acoustic mitigation infrastructure at scale.

Science Corporation restores reading in Stargardt patient with PRIMA chip

Source: techcrunch.com ↗

Neural interface startup Science Corporation placed its PRIMA retinal implant into a Stargardt disease patient for the first time, expanding beyond its existing European regulatory approval for geographic atrophy. The outpatient procedure was conducted at the Sydney Eye Hospital and took approximately one hour.

The hardware architecture uses a 2 mm by 2 mm subretinal photovoltaic array that replaces damaged photoreceptor cells. External glasses fitted with a camera capture the visual field and project near-infrared patterns onto the subretinal chip, which then stimulates the remaining functional inner retinal neurons. The patient was able to read individual letters during the first visual rehabilitation session and read complete words during the second. This demonstrates that subretinal optoelectronic prostheses can successfully bypass genetic retinal degeneration without requiring tethered intracranial brain implants.

Apple ships the $299 HomeView smart display

Source: nytimes.com ↗

Apple unveiled the HomeView, a 7-inch connected home display starting at $299. The dedicated home terminal comes in both a countertop format and a wall-mounted chassis, with shipments starting November 13.

The primary software hook is local identity switching: the device tracks and authenticates up to six household members, dynamically swapping interfaces, calendars, and automation controls based on who is standing in front of it. Instead of forcing full iPadOS overhead onto smart home hubs, Apple is shipping a purpose-built home dashboard targeted at smart home control and multi-user environments.

Behavioral harnesses turn cheap models into reliable coders

Source: martinfowler.com ↗

Writing on Martin Fowler’s blog, researchers published a practical benchmark studying the cost-benefit balance of behavioral harnesses for coding agents. The benchmark pitted the frontier GPT-6 Sol against the cheaper GPT-6 Luna on building an API service from scratch.

Running unharnessed, Luna succeeded in only 3 out of 10 test runs. Wrapped in an external behavioral test harness that enforces step-by-step verification and linting feedback, Luna hit 10 out of 10. Luna with a harness cost $0.273 per run, compared to $0.406 for Sol. The harness introduces significant overhead - roughly 10.5× higher token consumption and a 3× increase in execution time for Sol - showing that behavioral sandboxes are ineffective for frontier models but represent the most cost-efficient way to make smaller, cheaper models reliable in production CI pipelines.

DeployCos, vibe-coded apps, and dating market resets

Source: tanayj.com ↗

Enterprise software distribution is shifting toward high-touch models. Venture analysis details the sudden rise of DeployCos, spin-out service firms dedicated to bridging the diffusion gap between raw foundation models and complex legacy enterprise stacks. Anthropic backed its spin-out Ode with $1.5 billion in private equity support, OpenAI’s dedicated deployment arm has pulled in over $4 billion with a roster of over 100 Forward Deployed Engineers, and AWS committed $1 billion to its own forward-deployed unit. Enterprise AI adoption is currently bottlenecked not by API availability, but by integration labor.

Meanwhile, Google is altering distribution on mobile by letting AI Studio users export personal vibe-coded apps directly to the Play Store. The company is expanding Ask Play to build complex comparison charts across listings, alongside a new Game Finder tool powered by “Gamer DNA” profiles synthesized from 1.2 billion Android gaming profiles.

Finally, the consumer subscription model for dating took a massive blow: Bumble agreed to a $1.3 billion take-private acquisition by Ithaca Holdings. Facing a 16.4% plunge in paying subscribers (down to 3.2 million) and second-quarter revenue falling 15.2% year-over-year to $211 million, the platform is exiting public markets to completely overhaul the legacy swipe interface in favor of agentic matchmaking systems.