Today's brief — full text

OpenAI is paying AI-chip startup Cerebras Systems more than $20 billion over three years, Reuters reported April 17 — roughly double its previously disclosed commitment. The deal also grants OpenAI warrants for a minority stake in Cerebras. In an unusual twist, the arrangement is structured as a working-capital deposit: OpenAI is fronting Cerebras around $1 billion to build data centers and carrying that cash as an asset on its own balance sheet instead of expensing it. The move locks in a credible alternative to Nvidia and effectively solves the customer-concentration problem that derailed Cerebras's 2024 IPO. It also underscores the bigger story of 2026 — model leadership now lives or dies on access to chips, power, and data-center footprint, not just algorithms.

Anthropic released Claude Opus 4.7 on April 16, making its most capable public model generally available across Claude, the API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry. The update focuses on three fronts: software engineering, long-horizon agent work, and vision. Maximum image resolution jumps from 1,568px to 2,576px (3.75 megapixels), letting Opus read screenshots, diagrams, and documents far more reliably. A new effort parameter sits between high and max, giving developers finer control over the tradeoff between latency and depth of reasoning. Pricing stays put at $5 per million input tokens and $25 per million output — the same as Opus 4.6. Built-in safeguards automatically block requests linked to prohibited cybersecurity uses.

Anthropic overhauled the Claude Code desktop app on April 14 and launched Routines in research preview. The new Mac and Windows app adds a multi-session sidebar, drag-and-drop panes, an integrated terminal, and an in-app file editor with HTML and PDF preview. Routines are saved Claude Code automations that run on Anthropic's cloud, no laptop required. Configure a routine once — prompt, repository context, connectors — then trigger it by cron schedule, an HTTP call, or a GitHub event like a PR opening. Daily caps run 5 routines for Pro, 15 for Max, and 25 for Team/Enterprise, with more available for purchase. Together, the release nudges Claude Code from a chat tool toward infrastructure — a personal engineer that keeps working while you're asleep.

Google announced on April 16 that Gemini's Personal Intelligence feature can now generate images using real people and scenes pulled from a user's Google Photos library. Powered by Nano Banana, Google's image model, Gemini can reach into labeled groups of people and pets and place them inside custom pictures — ask for 'a claymation image of me and my family cooking' and Gemini produces it without a manual upload. The feature is opt-in and is rolling out to Plus, Pro, and Ultra subscribers in the US within days. Google says it does not train models on private photo libraries. The integration is the clearest signal yet that personal, signed-in data is becoming the main differentiator for AI assistants as the underlying foundation models commoditize.

Danish pharma giant Novo Nordisk announced an expansive partnership with OpenAI on April 14, deploying OpenAI models across drug discovery, clinical trials, manufacturing, supply chains, and commercial operations. The Ozempic maker wants to cut the time from research to patient, especially as rival Eli Lilly eats into its lead in obesity treatments. OpenAI will also help upskill Novo's global workforce on AI. Pilot programs start across R&D, manufacturing, and commercial in the coming months, with full integration targeted by year end. Financial terms were not disclosed. The deal is structured with strict data protections, human oversight, and governance guardrails — a nod to the regulatory scrutiny that comes with putting large language models anywhere near drug discovery and patient data.