Meta's Hyperion Eats a Parish; Cursor Ships Composer 2.5

Anthropic acquired Stainless, the SDK generator that has been turning API specs into production-ready client libraries for OpenAI, Google, Cloudflare, Replicate, and Runway since 2022. Anthropic confirmed it will wind down all hosted Stainless products. Customers keep what they have generated but lose the auto-update pipeline. The Information pegged the price above $300m. Founder Alex Rattray and the team move to Anthropic.

A federal jury in Oakland deliberated for under two hours before dismissing every claim in Elon Musk's $150 billion lawsuit against OpenAI, Sam Altman, and Greg Brockman. Judge Yvonne Gonzalez Rogers concurred. The decision rested on statute-of-limitations grounds, not the merits. Musk has confirmed he will appeal. The verdict removes the last major legal block on OpenAI's restructure and a possible IPO this year.

Meta told staff on Monday it is moving 7,000 workers into four new AI organisations with what HR head Janelle Gale called 'A.I. native design structures' and fewer managers per IC. Two days later Meta plans to cut roughly 8,000 employees, about 10% of headcount. The company has also folded AI usage into many performance reviews and is dialling back its metaverse spend to fund the shift.

Bloomberg Businessweek details Meta's Hyperion AI data centre in Richland Parish, Louisiana, now projected to cost Meta more than $200 billion across a campus that could cover much of Manhattan. The site will draw 5 gigawatts of power from ten new gas-fired turbines, run on $27 billion in Blue Owl debt kept off Meta's balance sheet, and was negotiated under NDA before locals knew it was happening.

Odyssey released Agora-1, a multi-agent world model that lets four players share a single GoldenEye-style deathmatch generated entirely by the model in real time. The architecture decouples shared world state from rendering, unlike Multiverse, Solaris, or MultiGen. Odyssey is pitching it as a 'learned game engine' with applications across gaming, robotics, defence, and education, and positions it as the first in a series of multi-agent worlds.

Composer 2.5 lands as Cursor's second iteration on Moonshot's open-source Kimi K2.5 checkpoint, with new targeted-RL training that injects textual feedback at the exact point in a rollout where the model went off the rails. Cursor says the result is steadier behaviour on long tasks and better instruction following. They have also started training a model 10x larger from scratch with SpaceXAI on Colossus 2.

Jason Liu describes pushing Codex beyond coding into knowledge work by building a durable operating loop around each project. He keeps pinned megathreads that he has been compacting for months, feeds them via Wispr Flow voice input so the agent sees unedited thinking, and treats memory and resumable tasks as first-class. The tradeoff is that revisited threads fall out of cache and cost more.

Cameron Wolfe publishes a long-form survey on evaluating agent systems built around the ReAct framework, manager and decentralised multi-agent setups, and context engineering patterns like dynamic context, summarisation, and note-taking. He argues static benchmarks no longer map onto how agents actually run and walks through recent benchmark case studies before laying out a roadmap for building your own eval harness.

Thariq Shihipar from Anthropic's Claude Code team argues HTML has overtaken Markdown as the right format for planning with agents. Plans now run thousands of lines, and HTML's interactive elements, mockups, and information density let humans actually engage rather than skim. Engineers, he says, are becoming 'compute allocators' who spend $500 of model time per task and need richer interfaces to stay in the loop.

Cloudflare's Grant Bourzikas pointed Anthropic's Mythos Preview at more than fifty of Cloudflare's repos under Project Glasswing. Two capabilities stood out: Mythos chains primitives like a use-after-free into a working ROP exploit, and it writes, compiles, and runs proof code in a scratch environment to confirm the bug. He calls the jump 'a different kind of tool doing a different kind of work.'

skids.dev works out the maths on Anthropic's 5-minute cache: a refresh ping costs 10% of the prefix's input price, so the breakeven against rewriting is 62.5 minutes of expected reuse. Below that, refresh. Above that, let it expire. The number holds across Opus, Sonnet, and Haiku and across prefix sizes from 5K to 500K tokens. The dollars change, the decision does not.

Alex Irpan argues that AI making things cheaper to produce never produces shorter working days. Textile automation drove demand for novelty T-shirts; machine guns did not shrink armies. With coding agents he finds himself with less downtime, not more, because if he is not running an agent on something, he is leaving capacity on the table. Writing code from scratch is for free time now.

Addy Osmani lays out cognitive debt: ship the fix without engaging with the problem and your mental model never updates. An Anthropic randomised trial had AI and manual groups finish Python library tasks at the same speed, but AI users scored 50% on a follow-up quiz versus 67% manual. The cut mattered: conceptual-question users hit 65%, copy-paste users under 40%.

Marc Brooker frames AI coding agents as feedback loops wrapped around a flawed open-loop component, the LLM. Drawing on his EE background, he shows feedback can turn multipliers into square rooters and planes into stable systems. The lens predicts which agent weaknesses are inherent and which are transient, which is the better question for deciding what to build now versus what to wait on.

Six months ago this newsletter called AI a bubble. The author concedes nothing has popped: Anthropic moved from $183B to talks at $900B, Big Four capex is projected at $725B for 2026 (1.8x year-on-year), and Google's market cap hit $4.8T. But the subsidy era is ending. All four major coding tools, Cursor, Copilot, Claude Code, and Codex, have abandoned flat-rate plans for metered, token-based pricing.

Tanay Jaipuria parses Unitree's $620 million Shanghai STAR Market IPO filing. Revenue jumped 335% from $58m in 2024 to about $252m expected in 2025, with gross margin near 60%. Unitree shipped roughly 5,500 humanoid units last year, against US competitors in the low hundreds. Around $300m of the proceeds is earmarked for VLA and World Model plus Action training over three years.

AirCaption is a Mac and Windows desktop app that transcribes and subtitles audio or video locally, with no internet and no upload limits. It supports 67 languages, uses OpenAI's latest speech models on-device, and keeps media private. Editing text and timing happens in-app, with hotkeys and folder-level batch transcription via drag-and-drop. The target users are video editors, podcasters, journalists, and researchers.

MinishLab released Semble, a code search library that returns exact snippets to agents using roughly 98% fewer tokens than grep plus read. It indexes a full codebase in under a second, runs on CPU with no API keys, and ships as an MCP server, a CLI, and an AGENTS.md snippet for Claude Code, Cursor, Codex, and OpenCode. Retrieval quality lands at 99% of a code-specialised transformer.

humanlayer's 12-factor-agents repo, a Heroku-style manifesto for building LLM-powered software that survives contact with production customers, crossed 20,000 stars on GitHub with 359 added on the day of release. The project is written in TypeScript and lays out twelve principles for keeping agents observable, resumable, and shippable. It is the closest thing the agent space has to a shared production checklist.