Overview
Three kinds of leverage shifted this week: OpenAI redirected compute and incentives toward enterprise distribution and advertising; agents gained more ways to act, increasing the value of permission and evaluation layers; and a court checked federal procurement power. Research and disclosure failures both favored systems that expose where consequential choices are made.
Developments
1. OpenAI reprices distribution and retreats from costly consumer video
OpenAI put three prices on distribution this week. Its proposed private-equity venture offered a 17.5 percent minimum return and early model access while seeking about $4 billion. Its US ChatGPT ad pilot passed a $100 million annualized run rate in six weeks. Meanwhile, the company ended Sora and an unclosed Disney partnership that contemplated a $1 billion investment.
The portfolio points away from compute-heavy consumer creation and toward channels that can subsidize acquisition or steer enterprise purchasing. Private-equity guarantees pay intermediaries to place OpenAI inside portfolio companies; ads monetize existing attention; Sora consumed scarce capacity without equivalent continuity. OpenAI gains distribution leverage, but its products become harder to compare independently of financing and placement incentives.
Sources: Reuters on OpenAI's private-equity proposal · Reuters on OpenAI ending Sora · Reuters on ChatGPT's advertising pilot
2. New action surfaces turn agent reliability into a permissions problem
Claude Code and Cowork added permission-gated mouse, browser, file, and developer-tool control, while Codex plugins began bundling instructions, integrations, and MCP servers into installable packages. OpenAI simultaneously opened a bounty for reproducible agentic prompt injection, exfiltration, and harmful actions; Anthropic described a three-session coding harness that separated planning, generation, and evaluation.
These releases move competition from answer quality to delegated authority. A plugin or visual-control session can cross authentication and data boundaries that text alone cannot, so permission scope captures downside before monitoring or a bounty can observe it. Independent evaluator sessions gain leverage because the system producing an action cannot also be the sole judge of completion.
Sources: Ars Technica on Claude computer use · OpenAI's Codex plugin documentation · OpenAI's Safety Bug Bounty · Anthropic on harness design for long-running development
3. Provenance failures spread from model lineage to shipped assets and behavior claims
Cursor initially omitted that Composer 2 began from Moonshot AI's Kimi K2.5, then said the base accounted for roughly one quarter of the final model's compute. Pearl Abyss began removing AI-generated Crimson Desert assets that it said were temporary. OpenAI described its public Model Spec as partly aspirational, often targeting behavior up to three months ahead of deployed models.
The three cases expose different gaps in the release record: inherited model dependencies, asset origin, and the distance between specified and observed behavior. Missing lineage transfers legal and audit work to customers; missing asset controls turn placeholders into production content; an aspirational specification can be mistaken for current capability. Disclosure after release leaves downstream users carrying uncertainty they did not create.
Sources: TechCrunch on Cursor's Kimi disclosure · Moonshot's Kimi K2.5 model card · The Verge on Crimson Desert's AI-art audit · OpenAI's approach to the Model Spec
4. Research shifts efficiency gains into the selector
PivotRL ran reinforcement-learning rollouts around intermediate turns where sampled actions produced sharply different outcomes, matching its agentic-coding baseline with four times fewer rollout turns. Perspective-Driven Inference directed scarce human labels toward demographic groups with weaker model estimates. TAPS routed generation through task-matched speculative-decoding drafters after naive checkpoint averaging diluted their specialization.
The common bet is selective expenditure: spend training compute, human judgment, or serving capacity where variation is consequential instead of treating every example or token alike. That moves leverage from the largest universal model to the component that detects pivots, underrepresented perspectives, or workload type. Efficiency is bought by making misclassification at the selector a new failure mode.
Sources: PivotRL preprint, version 1 · Multi-Perspective LLM Annotations preprint, version 1 · TAPS preprint, version 1
5. A federal court separates procurement choice from market exclusion
The Pentagon's court filing alleged that Anthropic posed a security risk; at a March 24 hearing, Judge Rita Lin questioned whether the resulting supply-chain designation was punitive and overbroad. On March 26, she granted a preliminary injunction against the federal ban, the contractor restriction, and the designation, finding Anthropic likely to prevail on retaliation, statutory, and procedural claims. The injunction itself was stayed for seven days.
The ruling preserves the government's freedom to stop using Claude while restricting its ability to turn one contract dispute into exclusion across agencies and private contractors. That boundary changes the parties' leverage: the Pentagon retains vendor choice, but broad procurement sanctions now carry an evidentiary and procedural burden before they can reshape an AI supplier's wider market.
Sources: Pentagon declaration in Anthropic v. Department of War · Wired on the March 24 hearing · District court order granting the preliminary injunction · Preliminary injunction terms