Wednesday 27 May 2026
xAI enters the coding agent race, Anthropic approaches profitability ahead of its October IPO, and the Pope weighs in on AI ethics. Meanwhile, DeepSeek’s permanent 75% price cut escalates the model pricing war, and the MCP protocol gets its biggest revision since launch.
🤖 Models and Launches
Introducing Grok Build — xAI launches a new coding agent and CLI in beta for SuperGrok and X Premium Plus subscribers, supporting plan mode reviews, headless mode, and parallel subagents.
Gemini 3.5 Flash Looks Good for How Fast It Is — Google’s latest Flash model outscores 3.1 Pro on Terminal-Bench and MCP Atlas while running 4x faster, positioning it as a daily driver for agentic workflows — though Zvi finds it unconvincing against Opus 4.7 or GPT-5.5 for non-latency-sensitive work.
GPT-5.6 Leaks: Coming in June — Leaked details suggest GPT-5.6 will focus on stronger multi-step reasoning, better agentic workflows, and improved frontend generation.
Anthropic Prepares Mythos 1 for Claude Code and Claude Security — Traces of Mythos have surfaced on Google Cloud and AWS through vulnerability discovery programs. Claude Opus 4.8 is also rumored in the works.
Lance: Unified Multimodal Model — ByteDance releases a 3B-parameter model supporting image/video understanding, generation, and editing, trained from scratch on a 128-A100-GPU budget.
🔬 Research and Engineering
On-Policy Distillation — Trains a student model on its own trajectories with teacher supervision via KL regularization, closing the train-inference distribution mismatch. Reverse-KL emerges as the default for mode-seeking smaller students.
Introducing BenchBench — A meta-benchmark testing how well models can create benchmarks. GPT 5.2 was the only winner; every other model from Opus 4.6 to GPT 5.5 struggled to create useful benchmarks.
AlphaProof Nexus Solves Decades-Old Math Problems — Google DeepMind’s system autonomously solved 9 of 353 open Erdős problems, including questions unanswered for decades, at inference costs of a few hundred dollars each.
The 2026-07-28 MCP Specification Release Candidate — The largest MCP revision since launch introduces a stateless core that scales on ordinary HTTP, extensions, OAuth-aligned authorization, and a formal deprecation policy. Ships July 28.
Evaluating Multi-Agent Systems at Scale — OpenAI outlines a macro-evaluation workflow for agentic systems that analyzes patterns across entire populations of traces rather than isolated failures.
🛠️ Agents and Tools
Microsoft Webwright — A SWE-style browser agent framework that achieves state-of-the-art results on long-horizon web tasks by giving agents a terminal to launch multiple browser sessions, capturing screenshots only when needed.
Agent Sandbox — Kubernetes SIG project for managing isolated, stateful singleton workloads with stable identity and persistent storage, ideal for AI agent runtimes.
Reasonix — A DeepSeek-native coding agent engineered around prefix-cache stability, designed to be left running with low token costs across long sessions.
Auth.md — A new standard for secure agent registration: a Markdown file that tells agents how to register on behalf of users, with identity providers vouching for users without human interaction.
💰 Funding and Business
Anthropic’s March to Profitability — On track for $10.9B in Q2 revenue (up from $4.8B Q1), with Claude Code alone at $2.5B. Compute cost per revenue dollar dropped from 71 cents to 56 cents. Expects $559M profit by its October IPO.
DeepSeek Made Its 75% Discount Permanent — V4 Pro prices now sit below GPT-5, Claude Opus 4.7, and Gemini 3.5 Flash, with the widest gap against frontier reasoning models. The AI price war escalates.
The Neocloud Boom — SpaceX acquired xAI and is renting Colossus data centers to Anthropic for $15B/year. The broader AI buildout adds up to $7.5T of spend over 4.5 years — on par with the 1880s railroad boom.
Microsoft’s Quiet Claude Code Retreat — Microsoft is canceling most direct Claude Code licenses in its Experiences and Devices group, telling engineers to migrate to GitHub Copilot CLI by June 30. Enterprise AI coding unit economics don’t work at current token prices.
🔒 Security
A Hacker Group Poisoning Open Source at Unprecedented Scale — TeamPCP has corrupted hundreds of open source tools and breached GitHub via a poisoned VSCode extension. The financially motivated group deploys ransomware and sells victim data.
Bumblebee Goes Open Source — Perplexity open-sources a read-only security scanner that identifies risky packages, extensions, and AI tool configurations on developer machines.
Measuring LLMs’ Ability to Develop Exploits — Mythos Preview can turn vulnerabilities into exploit primitives and chain them into complete attack sequences, consistently outperforming all other models on ExploitBench and ExploitGym.
🌐 Notable
Musk and Altman’s AI Rivalry Reaches Boiling Point — As SpaceX and OpenAI race toward IPOs, a tiny circle of tech leaders tightens its grip on AI’s future.
Notes on Pope Leo XIV’s Encyclical on AI — The Pope’s new document covers environmental impact, algorithmic decision-making risks, power concentration in private actors, and compares AI to the Tower of Babel. Approachable even to non-Catholics.
US Students Boo Pro-AI Graduation Speakers — Recent grads are not fond of commencement speakers hyping technology they see as a threat to their career prospects.
NASA Selects Blue Origin for First Lunar Mission — Three lunar landings planned this year in preparation for a $20B moon base.
🔥 Hacker News
Using AI to Write Better Code More Slowly (415 comments) — Many think the point of AI coding is to write low-quality code fast, but it can also be used to write high-quality code.
Spain Blocks Prediction Markets Polymarket, Kalshi (313 comments) — Spain blocks the platforms over lack of gambling licences.
Dropbox CEO Drew Houston to Step Down (263 comments) — Houston steps down as Dropbox CEO.
The User Is Visibly Frustrated (232 comments) — A piece on the experience of dealing with frustrating software.
Outsourcing Plus Local AI Will Soon Become More Economical vs. Frontier Labs (231 comments) — The economics of local AI are shifting rapidly.
Uber President Says AI Spending Getting Harder to Justify (128 comments) — Uber’s president questions the ROI on AI investment.
🌍 World and Australia
Seven Heat Deaths in France as May Records Set Across Europe — Extreme early-summer heat brings highest May temperatures to UK, Ireland, and western Europe.
Iran Remains in Peace Talks Despite First US Strikes Since Ceasefire — Tehran condemns “definitive violation” but announces no specific reprisals as negotiations near decisive stage.
Quad Countries Ink Deal to Build Surveillance Network and Fiji Port — Australia, US, Japan, and India agree on infrastructure to counter China’s Pacific influence.
Moscow Wants to Destabilise Europe, EU Chief Warns — Von der Leyen visits Lithuania amid drone incursions as countries summon Russian ambassadors over Kyiv threats.
Sources: TLDR (General + AI), The Guardian, Hacker News — 4 newsletters processed, ~40 stories distilled