Overview
The gap between impressive capability and sustainable deployment is unusually visible today. OpenAI is ending Sora despite a major partnership, Anthropic's government dispute is testing procurement power, Claude can now operate a desktop, and Anthropic's own research emphasizes the value of experienced users.
Developments
1. OpenAI shuts down Sora and unwinds its Disney deal
OpenAI publicly announced that it was ending Sora. Reuters reported that the decision also ended a proposed three-year Disney partnership involving a $1 billion investment and more than 200 characters; the transaction had not closed and no money had changed hands.
Reuters cited Sora's compute demands and OpenAI's shift toward coding and enterprise products. Customers need a dated shutdown schedule for the app and API, export tools for existing work, and clarity on licensed assets. A prominent launch and strategic partner did not guarantee product continuity.
Sources: Reuters on OpenAI ending Sora · Ars Technica on the Sora shutdown
2. A judge questions the breadth of the Pentagon's Anthropic action
At a March 24 hearing, US district judge Rita Lin said the Pentagon's supply-chain-risk designation appeared punitive and insufficiently tailored to its stated security concerns. Anthropic was seeking a temporary pause; the judge's comments indicated skepticism, not a ruling at that point.
The distinction matters because the designation could affect contractors' commercial use beyond the disputed military work. Procurement restrictions should have a documented nexus to the risk, a bounded scope, and a review process, especially when the government can reshape a vendor's private market before a case is decided.
Sources: The Wall Street Journal on the Anthropic hearing · Wired on the Anthropic hearing
3. Claude Code and Cowork gain direct computer use
Anthropic added a research-preview computer-use mode to Claude Code and Cowork for Pro and Max subscribers on macOS. The tools can ask permission to point, click, browse, open files, and run developer tools when a connector is unavailable; Anthropic says this path is slower and more error-prone than connectors.
Anthropic also warns that safeguards are not absolute and that the model can see sensitive material displayed on screen. A defensible trial starts in isolated accounts, excludes financial and production systems, requires action-level approval, and uses scoped APIs instead of visual control wherever possible.
Sources: Ars Technica on Claude computer use
4. Vibe Physics is a practitioner account, not an autonomous discovery claim
Anthropic published physicist Matthew Schwartz's guest account of using Claude as a graduate-student-like collaborator in theoretical work. The article describes productive assistance while repeatedly emphasizing that domain expertise was necessary to frame questions, detect mistakes, and decide which paths were meaningful.
That is a more useful deployment model than replacing the scientist. Agents can expand search and calculation, while a defensible discovery claim includes traceable derivations, independent checks, and an expert able to recognize a plausible but physically invalid result.
Sources: Anthropic's Vibe Physics guest article
5. Anthropic finds an association between tenure and successful use
Anthropic's Economic Index analyzed Claude usage from February 5 through 12. After controls, users with at least six months of tenure had a roughly four-percentage-point higher conversation success rate; the report also found that the ten most common Claude.ai tasks fell from 24 to 19 percent of traffic since November.
Anthropic explicitly notes cohort and survivorship bias: early or continuing users may differ from newer users. The association supports investment in workflow training and repeated practice, while leaving both causality and the applicability of Claude-user behavior to the wider economy unresolved.
Sources: Anthropic Economic Index: Learning curves
6. Anthropic separates generation from evaluation in long coding runs
Anthropic described a three-agent harness that assigned planning, generation, and evaluation to separate Claude sessions. Its evaluator used Playwright against running applications, while structured contracts and artifacts defined completion criteria; one full-stack experiment ran about six hours and cost roughly $200.
The examples are internal case studies rather than controlled proof that more agents always improve software. They identify a practical tradeoff: independent evaluation found functional gaps that self-review missed, but the extra loop added substantial latency and cost, making selective use more defensible than universal orchestration.
Sources: Anthropic on harness design for long-running development