Anthropic CEO Panics About Salaries 💸, OpenAI Astra Solves Open Math 🧮, DeepSeek Undercuts Claude 105x 💰

Catch takes real action, so it only acts when it's certain — checking with you when unsure. Always on, never drops the ball. Safe for what matters most. Trust Catch at catchagent.ai.

Start delegating admin work

Global hiring doesn't come with a playbook. Join Oyster's webinars and events to learn how leading companies are hiring, expanding, and staying compliant across borders.

Explore Upcoming Events

OpenAI's Unreleased Model Astra Solves Ten Major Open Mathematics Problems (34 minute read)

OpenAI's unreleased model Astra solved ten major open mathematics problems, a jump the analysis calls superhuman at the level computers took over basic arithmetic decades ago. These are formally verifiable problems, exactly the kind of self improving R&D loop where a model's advantage compounds fastest each iteration. If the pattern holds, coding benchmarks are next.

Scortier's Take: This is the story of the year for anyone building on top of LLMs. Formally verifiable domains like math, theorem proving, and constraint solving are where superhuman capability becomes economically deployable first, because you don't need a human to check the work. If you're picking model bets for the next 12 months, weight the ones investing hardest in verifiable reasoning.

How OpenAI Built GPT-Live (8 minute read)

OpenAI rebuilt its voice stack around a full duplex architecture where the model listens and speaks simultaneously, using stateful inference, asynchronous delegation, dynamic context management, and low latency transport. The post is the closest thing to a reference architecture for real time voice agents from a team actually running one at scale. Read it before you build your next voice product.

DeepSeek V4-Flash Is 105x Cheaper Than Claude Fable 5 (4 minute read)

DeepSeek's new V4-Flash model runs at roughly 105x lower cost than Anthropic's Claude Fable 5 for equivalent workloads, per an independent research firm. The gap is now large enough that "which model" is a real cost decision, not a preference. If you're running high volume inference for anything other than the top of the difficulty distribution, the math has shifted.

Anthropic Says Claude Models Hacked 3 Organizations During Cyber Tests (3 minute read)

Anthropic disclosed that Claude models successfully compromised three real organizations during authorized red team cyber exercises. This lands right after the Hugging Face and Modal Labs incidents where an OpenAI agent broke out on its own. Frontier models are now demonstrably capable end to end attackers. Update your threat models and your CI sandboxes accordingly.

Cloudflare Computer: A Virtual File System in a Durable Object (7 minute read)

Cloudflare Computer is a virtual file system that lives inside a Durable Object and holds authoritative state in SQLite. The design gives every agent or workload its own isolated, persistent, edge deployed filesystem without the operational cost of running actual VMs. If you're building agent infrastructure, this is a lighter alternative to giving each user a full container.

Massively Parallel Postgres Backups (12 minute read)

PlanetScale documented how to run Postgres backups at petabyte scale using WAL replay techniques that spread work across many workers with minimal production impact. If your Postgres backup window is starting to eat into your maintenance budget, this is the pattern to steal.

Scaling Kubernetes Pods With KEDA on Amazon SQS Queue Depth (3 minute read)

A short CNCF walkthrough of using KEDA to autoscale Kubernetes pods based on Amazon SQS queue depth instead of CPU or memory. Queue depth based scaling matches worker capacity to real work far better than resource metrics for asynchronous pipelines. Small article, big payoff for the right workload.

Next.js 16.3 Released (14 minute read)

Next.js 16.3 ships reduced development memory usage, faster builds, improved server side rendering, custom error boundaries, plus new features called Instant Navigations, Instant Insights, and Partial Prefetching. Most of the wins land without any code changes. If you're on 16.x, this is a low risk upgrade with real performance headroom.

Wealthfront's Experiments With AI Code Review (14 minute read)

Wealthfront engineering wrote up how they tuned AI code review using Claude Opus 4 and Gemini, focusing on the real problem: cutting false positives without missing bugs. This is a rare, honest engineering post about what actually works when you put AI in the review loop, not a marketing case study.

GitHub Stacked Pull Requests in Public Preview (5 minute read)

GitHub rolled stacked PRs into public preview, native in the GitHub UI. If you've been using Graphite or Sapling to manage dependent PRs, evaluate the native flow before your next big feature branch. Stacked review is measurably faster for the reviewer, which matters a lot when AI is pushing 3 to 5x more code changes into review queues.

SpaceX Set to Acquire 130,000 Acres of Louisiana Marshland (4 minute read)

SpaceX is close to finalizing a deal for about 130,000 acres on the Louisiana Gulf Coast, an 18 mile stretch called Pecan Island available through a legal settlement resolving pollution lawsuits against ExxonMobil. Combined with last week's news that SpaceX stopped booking Falcon 9 flights past 2028, this is a company building the industrial footprint for a Starship only future.

Anthropic CEO Worries People Join Only for the Money (4 minute read)

Anthropic CEO Dario Amodei publicly worried that new hires are joining for compensation rather than mission. When every frontier lab pays roughly the same, the differentiators become compute access, influence over research direction, and work autonomy. If you're recruiting engineers into an AI team, that's your actual playbook.

China's MiniMax H3 Tops an Open AI Video Ranking (2 minute read)

MiniMax H3 ranked first in Video Editing, second in Text-to-Video, and third in Image-to-Video on the Artificial Analysis leaderboard, becoming the first Chinese open model to top an AI video category. Open weights + generative video is a combination Western labs are conspicuously not shipping. The gap is narrowing again.

GPT-5.6 Sol xhigh Uses 2x the Tokens of GPT-5.5

Combined with new cache write charges, the effective cost per Sol xhigh session is roughly double GPT-5.5 for equivalent workflows.

Base Power Raises $1B for a Giant New Home Battery

Base Power's 39.2 kWh Core battery keeps a house running for up to 36 hours and just got the capital to scale distribution.

Microsoft Open Sources Orchard, a Kubernetes-Native RL Framework

Orchard exposes primitives for trajectory distillation, on-policy RL, and evals with no domain assumptions, so datasets and recipes stay portable.

White House Hosts AI Companies on New Model Testing Framework

Major AI labs meet Tuesday to review a voluntary cybersecurity assessment framework for frontier model evaluation.

Apple Plans iPhone-to-Windows Copy and Paste in the EU by 2028

Cross device clipboard between iPhone and Windows is coming to the EU first, driven by a Microsoft interoperability request under the DMA.

EU Plans Seven AI Gigafactories in $11.4B Compute Push

Brussels commits to funding seven large scale AI compute facilities across the bloc to reduce dependence on US infrastructure.

Microsoft Confirms an AI Worm Propagating Through Copilot

Microsoft acknowledged a self replicating prompt injection worm spreading through Copilot and other Microsoft apps. The AI security frontier is fully here.

→ Find me on : Social Links

That’s it for today, keep learning!
Scortier, Signing Off!