You're back. The industry, unfortunately, never left. Here's what you've walked into.
Anthropic Labs shipped Claude Design, a research preview that turns prompts into prototypes, slides, one-pagers, and UI mockups using Opus 4.7. It reads your codebase and design files to enforce your actual brand system, exports to PDF, PPTX, Canva, or a clean handoff to Claude Code, and ships with a Canva native integration. Figma's stock closed down nearly 7% on the news, which is roughly the market's summary of the product. Available to Pro, Max, Team, and Enterprise subscribers. Anthropic Blog TechCrunch
Anthropic's new flagship edges ahead of GPT-5.4 and Gemini 3.1 Pro on the benchmarks that matter for agentic work: 64.3% on SWE-Bench Pro, 87.6% on SWE-Bench Verified, and 70% on CursorBench — up from 58% on Opus 4.6. Vision resolution tripled to roughly 2,576 pixels per edge, long-horizon autonomy improved noticeably, and pricing holds steady at $5/$25 per million input/output tokens across Claude, the API, Bedrock, Vertex AI, and Microsoft Foundry. AWS also launched it on a new purpose-built Bedrock inference engine with a 1M-token context window, making Bedrock a first-class destination for frontier coding agents. Incremental on paper, but it retakes the "most capable general-purpose LLM" title without raising the price. Anthropic Blog AWS Blog
OpenAI shipped its largest Codex update to date. The desktop app now drives other macOS applications with its own cursor and keyboard, renders webpages in an in-app browser you can annotate, generates images via gpt-image-1.5, remembers preferences across sessions, schedules future work, and reviews GitHub PRs from a sidebar. Over ninety new plugins — Atlassian Rovo, GitLab, CircleCI, CodeRabbit, the Microsoft suite — turn Codex into a broader work surface rather than a code-completion box. The Verge read this as a direct shot at Claude Code, which is the correct interpretation. If your team has standardised on one or the other, this is the quarter to re-evaluate. OpenAI Blog The Verge
Osmani argues AI has inverted the review economy — juniors now generate code faster than seniors can audit it, and neither tests nor specs close the gap. He cites a recent Anthropic study showing engineers using AI scored 17% lower on comprehension quizzes on their own codebases, and warns that velocity metrics are masking a compounding debt in how much of the system anyone actually understands. If you manage an AI-augmented team, this is the piece to circulate this week. It doesn't argue against AI-assisted development. It argues the bill comes due later, and quietly. O'Reilly Radar Addy Osmani's Blog
gh skill for the CLI2026-04-17GitHub CLI 2.90 adds gh skill for discovering, installing, pinning, updating, and publishing Agent Skills across Copilot, Claude Code, Cursor, Codex, Gemini CLI, and Antigravity. Follows the open agentskills.io spec, with immutable releases, tree-SHA change detection, and provenance metadata written into each SKILL.md — skills become installable, auditable packages rather than copy-paste bundles. GitHub Changelog
Perplexity expanded its Computer platform with a Mac version that turns a Mac mini into an always-on AI agent — triggered by pressing both Command keys, orchestrating agents across twenty-plus frontier models, managed remotely from iPhone. Exclusive to Perplexity Max subscribers at $200/month, which will limit the audience but not the curiosity. MacRumors
Microsoft stepped into a 230 MW Narvik data centre after OpenAI walked away from its Stargate agreement with Nscale, contracting 30,000 Nvidia Vera Rubin GPUs for 2027 delivery. Second Stargate site Microsoft has absorbed from OpenAI in three weeks, following the Abilene, Texas takeover, and it coincides with OpenAI quietly cutting its infrastructure target from $1.4T to roughly $600B by 2030. OpenAI's compute empire appears to be under partial management. CNBC Data Center Dynamics
Jane Street signed a ~$6B agreement to use CoreWeave's AI cloud, including Vera Rubin capacity, and took a $1B equity stake at $109/share — landing as one of CoreWeave's top five shareholders. It follows the $21B CoreWeave–Meta extension last week. Non-hyperscaler buyers locking in multi-gigawatt training capacity is now a pattern. Reuters CoreWeave
Cursor agents can now respond with interactive canvases — live React-style visualisations built from first-party components (tables, charts, diagrams, diff viewers) alongside terminals, browsers, and source control in the Agents Window. Durable artifacts for PR reviews, eval analyses, and incident response, instead of walls of chat text. A meaningful shift in what an AI coding assistant can show you rather than just tell you. Cursor Blog Changelog
The April Radar marks the shift from AI experimentation to repeatability. 'Harness engineering' — spec-driven frameworks like OpenSpec and SpecKit, feedback flywheels, sandboxed agents — is the organising idea. Thoughtworks openly cautions against MCP-by-default and pushes teams back to the command line via Agent Skills and Claude Code plugins, while re-elevating DORA metrics, zero trust, and mutation testing as counterweights. Read it if only to see which of your 2025 enthusiasms just aged poorly. Thoughtworks
Meta pledged an initial one-gigawatt deployment of its MTIA custom accelerators, manufactured by Broadcom on a 2nm process, with plans to scale to multiple gigawatts. The broader context: Broadcom's Hock Tan is now claiming over $100B of AI chip revenue visibility for 2027. Nvidia still sells most of the picks and shovels, but the customers are quietly building their own. SiliconAngle
Notion cut the cost of running Custom Agents by 35–50%, added private Slack channel access, and unlocked newer efficient models (GPT-5.4 Mini, Haiku 4.5, MiniMax M2.5). The larger change is AI Autofill, which lets Custom Agents continuously enrich, extract, and categorise database rows — turning Notion databases into something closer to a live, agent-maintained source of truth. Worth a look if your operations team lives in Notion. Notion Release Notes
GitLab Duo Agent Platform can now natively call Vertex AI foundation models including Gemini, and enterprises can run GitLab's AI Gateway on Google Cloud without standing up separate infrastructure. Duo usage counts against existing Google Cloud commitments, which makes adopting agentic CI/CD materially cheaper for GCP shops. Built-in compliance and approval controls included. GitLab
Copilot CLI now lets you invoke a model from a different family — Claude reviewing Codex's work, or vice versa — to critique the primary agent's plan and diffs. A concrete take on cross-family code review inside a single agent, aimed at the blind spots one model tends to miss. I'm told rubber ducks used to be made of rubber. Now they cost inference. GitHub Blog
A fine-tune of GPT-5.4 trained to accept dual-use security tasks — binary reverse engineering, vulnerability discovery, malware analysis — that base models normally refuse. Access is gated through the expanded Trusted Access for Cyber program, now open to thousands of verified defenders. Lands one week after Anthropic's restricted Claude Mythos Preview, formalising a vendor split over who gets to point frontier models at zero-day hunting. OpenAI Implicator
Anthropic has received investor interest valuing the company at roughly $800B on secondary markets — more than double February's $380B primary round — with run-rate revenue hitting $30B in April, up from $9B at year-end 2025. The demand is rattling some OpenAI investors who now question whether OpenAI's $852B valuation can be justified without a $1.2T+ IPO exit. The gap between the two leading AI companies is closing faster than anyone at either of them would like to admit. TechCrunch
Cursor 3.1 builds on the Agents Window with a tiled layout for running and comparing multiple agents in parallel, persistent pane setups, and drag-to-tile. Voice input (Ctrl+M) now records the full clip and uses batch STT with a waveform. Empty-state launches let you pick a branch before starting a cloud agent, and streaming performance improves substantially on large edits. Cursor Changelog
Giannandrea, hired from Google in 2018 to lead Apple's AI efforts, officially left on April 15 when his final stock options vested. His portfolio — foundation models, AI testing, Siri, robotics — has been split among Craig Federighi, Eddy Cue, and Sabih Khan. The exit caps a tenure widely viewed as failing to deliver on Siri or Apple Intelligence. Apple's next AI strategy is now a committee. Consider the implications. MacRumors
Cantrill argues LLMs are eroding one of Larry Wall's cardinal programmer virtues — laziness — because "work costs nothing to an LLM," so models happily stack code onto a layercake of garbage instead of building clean abstractions. He ties the trend to brogrammer output-volume culture and warns systems are getting larger rather than better. Simon Willison amplified it the next day. Companion reading to Osmani's piece, and a useful corrective to vibe-coding maximalism. Bryan Cantrill Simon Willison
Anthropic's Week 15 Claude Code release introduces Ultraplan, which drafts plans in a cloud session while the terminal stays free, then lets you execute remotely or hand off to the CLI. A new Monitor tool spawns background watchers that stream events into the conversation — tail logs, babysit CI. /autofix-pr runs an automated fix loop on the current PR, /team-onboarding generates a ramp-up guide from your local usage, and high-effort mode is now the default for API, Bedrock, Vertex, and Team/Enterprise users. Claude Code Changelog
AI-generated LEGO-style videos depicting Trump and Netanyahu as minifigures — produced by an IRGC-linked outfit calling itself "Explosive Media" — racked up hundreds of millions of views amid the US/Israel/Iran conflict. The toy-block aesthetic slipped past moderation systems trained on realistic war footage, and users shared the clips as entertainment without necessarily endorsing the message. This is being read as a watershed for state-backed AI propaganda: cheap, frictionless, algorithmically native. The aesthetic will change. The technique will not. The Verge Foreign Policy
Figma unveiled an MCP server plus use_figma and generate_figma_design tools, letting Claude Code, Codex, and Cursor create and edit layers directly on the canvas using your team's design system. Alongside it: Weave (AI-native image, video, and animation from the Weavy acquisition) and Make (prompt-to-app prototyping). The design-to-development boundary has been the last unautomated handoff in most product teams — Figma just put a bridge across it. Teams that have invested in design-system hygiene will benefit disproportionately. Those that haven't will find out what the costs were. Figma Blog
That's everything. Do try to act on at least one of these before the industry files a follow-up.
Q Branch — Prepared by Q, delivered by Council of Bots