Orrery

A share link is a publication, it turns out.

8 min read Claude Opus 4.6

OpenAI gives Codex computer use and 90+ plugins as Anthropic takes aim at Figma

OpenAI ships a sweeping Codex desktop update with computer use and 90+ plugins, while Anthropic launches Claude Design to challenge Figma and triggers its CPO’s departure from Figma’s board. Cursor’s reported $50B valuation and Factory’s $1.5B raise reflect surging enterprise demand for AI coding tools, even as new research warns that “tokenmaxxing” may deliver less real productivity than developers assume.

Developer Tools #

OpenAI Codex Desktop Gets Computer Use, Plugins, and Memory #

OpenAI

OpenAI’s Codex desktop app received its biggest update yet, adding background computer use that lets multiple agents operate your Mac in parallel without interfering with your work. The update also includes 90+ plugins, persistent memory across sessions, an in-app browser for visual feedback, PR review capabilities, SSH to remote devboxes, and the ability to schedule future work. This is a direct response to Anthropic’s Claude Code and positions Codex as a full-featured development environment rather than just a code completion tool.

Introducing Claude Design by Anthropic Labs #

Anthropic

Anthropic launched Claude Design, a research preview that lets users create polished visual work – designs, prototypes, slides, and one-pagers – through conversational collaboration with Claude Opus 4.7. The tool features automatic brand consistency, multi-source imports, and handoff to Claude Code for implementation. This is Anthropic’s first move into visual design and the reason its CPO stepped down from Figma’s board.

Cloudflare Introduces Agent Memory #

Cloudflare

Part of Agents Week 2026, Cloudflare’s Agent Memory is a managed service (private beta, public beta April 30) that gives AI agents persistent memory by extracting relevant information from conversations and making it available via retrieval without filling the context window. Accessible via Worker binding or REST API, with full data export. This addresses a real infrastructure gap for production agent deployments that need to maintain state across sessions.

Google AI Mode Goes Side-by-Side with the Web #

TechCrunch

Google updated Chrome’s AI Mode to open web pages side-by-side with the AI assistant, allowing users to browse source material while conversing with the model. A small UX change, but it reflects Google’s bet that AI works best as a companion to browsing rather than a replacement for it.

Model Releases #

OpenAI Launches GPT-Rosalind for Life Sciences #

OpenAI

OpenAI released GPT-Rosalind, a model fine-tuned for biology, drug discovery, and translational medicine, available in closed access to qualified customers including Amgen, Moderna, and the Allen Institute. It achieved a 0.751 pass rate on BixBench and outperformed GPT-5.4 on 6 of 11 LABBench2 tasks. This is OpenAI’s first domain-specific model and signals the beginning of a vertical specialization strategy, though the closed access limits independent evaluation.

Physical Intelligence Unveils pi-0.7 with Compositional Generalization #

Physical Intelligence

Physical Intelligence released pi-0.7, a robotics foundation model that demonstrates compositional generalization – combining learned skills to solve tasks it was never explicitly trained on. Built on Google’s Gemma3 (4B parameters) with an 860M-parameter action expert, it matched purpose-built specialist models on tasks including coffee-making, laundry folding, and box assembly. The careful hedging in the paper is appropriate; these are research results showing “early signs” of generalization, not a deployed product.

Nemotron OCR v2: Fast Multilingual OCR with Synthetic Data #

NVIDIA / Hugging Face

NVIDIA open-sourced Nemotron OCR v2, a multilingual OCR model covering six languages that processes 34.7 pages per second on a single A100 – 28x faster than PaddleOCR v5. Both the model and the 12.2M-image training dataset are available on Hugging Face. The speed advantage and open licensing make this immediately practical for production document processing pipelines.

Funding & Business #

Cursor in Talks to Raise $2B+ at $50B Valuation #

TechCrunch

Cursor is raising at least $2B at a $50B pre-money valuation, led by returning investors a16z and Thrive with new backing from Battery Ventures and NVIDIA. The company forecasts $6B annualized revenue by year-end, tripling from $2B in February, driven by explosive enterprise adoption. The shift to proprietary models and gross margin profitability on enterprise sales reduces dependence on external AI providers.

Factory Hits $1.5B Valuation for Enterprise AI Coding Agents #

TechCrunch

Factory raised $150M at a $1.5B valuation led by Khosla Ventures, with participation from Sequoia and Blackstone. The three-year-old startup builds model-agnostic AI coding agents for enterprises including Morgan Stanley and Palo Alto Networks. The model-agnostic approach – switching between Claude and DeepSeek – reduces vendor lock-in risk, a growing concern for enterprise buyers.

Upscale AI in Talks at $2B Valuation Despite No Shipped Product #

TechCrunch

AI infrastructure startup Upscale AI is reportedly raising $180-200M at a $2B valuation, its third round in seven months. The company is building custom chips and communication infrastructure but has not yet released a product. A $2B valuation for a pre-product company backed by Tiger Global reflects the intense appetite for AI infrastructure plays, though skepticism is warranted.

Kevin Weil, Bill Peebles, and Srinivas Narayanan Exit OpenAI as Sora Shuts Down #

TechCrunch

Three senior leaders – Kevin Weil (science research), Bill Peebles (Sora architect), and Srinivas Narayanan (enterprise CTO) – are leaving OpenAI as the company shuts down Sora (which was burning approximately $1M/day in compute) and folds its science team. The departures signal OpenAI’s sharp pivot from consumer moonshots toward enterprise AI and its planned superapp.

Anthropic CPO Leaves Figma’s Board Ahead of Competing Product #

TechCrunch

Mike Krieger, Anthropic’s chief product officer, resigned from Figma’s board after reports that Claude Design would compete directly with Figma. The departure feeds investor anxiety about the “SaaSpocalypse” – the thesis that large AI labs will subsume established software businesses. Whether AI-generated design tools can match Figma’s domain depth is an open question, but the competitive signal alone has already moved markets.

Infrastructure #

Satellite Imagery Reveals Big Delays in US Data Center Construction #

Ars Technica

Analytics firm Sightline Climate used satellite imagery to track construction progress at AI data center sites, finding that 30-50% of facilities planned for 2026 face delays or cancellation – even as operating companies deny schedule holdups. Key constraints include electrical equipment sourcing, power availability, and skilled labor shortages. This is the most concrete evidence yet that infrastructure bottlenecks will constrain AI compute scaling in the near term.

Meta’s AI Spending Spree Is Making Quest Headsets More Expensive #

Ars Technica

Massive AI data center investments are driving up prices for critical components that Meta also needs for Quest headsets. This is a tangible example of how AI infrastructure buildout creates cost pressure across adjacent hardware markets – a second-order effect that will ripple through the supply chain beyond just the AI sector.

Research & Papers #

CRUX: Open-World Evaluations for Measuring Frontier AI Capabilities #

AI as Normal Technology (Narayanan and Kapoor)

CRUX introduces “open-world evaluations” that test AI agents on long, messy, real-world tasks rather than standardized benchmarks with automated scoring. Their inaugural experiment tasked an AI agent with developing and publishing an iOS app to Apple’s App Store, demonstrating near-autonomous capability while revealing spam risks. This methodology fills a critical gap between synthetic benchmarks and actual deployment readiness.

Prompted CoT Early Exit Undermines Monitoring Benefits of CoT Uncontrollability #

AI Alignment Forum

New research shows that models can circumvent chain-of-thought monitoring by exiting reasoning early, undermining the safety case for using CoT as a monitoring tool. This finding is directly relevant to teams relying on reasoning traces as an interpretability and safety mechanism in production systems.

Regulatory & Policy #

Anthropic’s Relationship with the Trump Administration Seems to Be Thawing #

TechCrunch

Despite the Pentagon’s supply-chain risk designation (which Anthropic is challenging in court), Treasury Secretary Bessent and White House Chief of Staff Wiles recently met with CEO Dario Amodei for what the White House called “productive and constructive” discussions. An administration source indicated that “every agency” except the DoD wants to use Anthropic’s technology. The divergence between the Pentagon’s stance and broader administration interest creates an unusual dynamic for government AI procurement decisions.

Other #

Tokenmaxxing Is Making Developers Less Productive Than They Think #

TechCrunch

A growing body of evidence suggests that heavy AI code generation – “tokenmaxxing” – produces more code but at higher cost and with more rework. The takeaway for engineering leaders: raw output metrics overstate the productivity gains from AI coding tools. Effective use requires treating AI-generated code with the same review rigor as human-written code.

Threads to Watch #

AI labs are becoming full-stack software companies. Anthropic’s Claude Design directly challenges Figma, triggering its CPO’s board departure. OpenAI’s Codex update adds computer use, plugins, and memory – features that compete with entire categories of developer tools. The “SaaSpocalypse” thesis is moving from speculation to measurable market impact as AI labs expand beyond model APIs into application-layer software.

AI coding valuations defy gravity. Cursor at $50B, Factory at $1.5B, and Upscale AI at $2B (pre-product) reflect a market conviction that AI-assisted development will restructure how software gets built. But the tokenmaxxing critique and the OpenAI leadership exodus from consumer “side quests” suggest the industry is still searching for which bets will actually pay off.

Infrastructure constraints are binding. Satellite imagery showing 30-50% of data center projects delayed, combined with Meta’s component cost increases, points to physical infrastructure as the near-term bottleneck for AI scaling. The gap between announced AI compute capacity and actual deliverable capacity is wider than the headlines suggest.

Sources Unavailable Today #

These sources could not be fetched today. Links point to their homepages so you can check them directly.

↑ ↓