March

8 events

Editors' summary

GPT-5.4 puts computer use in the general model, and Sora folds up

On 5 March, OpenAI released GPT-5.4, the first of its general-purpose models with computer use built in — reading the screen and driving mouse and keyboard. It scored 75.0% on OSWorld-Verified, a benchmark of desktop tasks, which OpenAI said put it above the human score. Smaller mini and nano models followed on the 17th, with mini reaching free accounts.

On 24 March, OpenAI announced it was shutting down Sora, its video generation social app, less than six months after opening it on 30 September 2025. The app and web version would stop on 26 April, the API on 24 September.

Reports put it down to inference costs the revenue did not cover, with usage falling. OpenAI said it was redirecting compute to coding and enterprise work, and would continue video generation as world model research.

The vocabulary on the user side grew too. "Comprehension debt" — AI-generated code piling up that no one on the team understands — began circulating among developers around this time, a coinage modeled on technical debt.

This block is written by the editors. It is kept separate from the sourced record below.

Record8 events

5
ModelOpenAI
GPT-5.4 released, with computer use built into the general model
OpenAI released GPT-5.4, the first of its general-purpose models to have computer use built in — reading the screen as images and driving mouse and keyboard. It scored 75.0% on OSWorld-Verified, a benchmark of desktop tasks, well up from GPT-5.2's 47.3%. OpenAI described this as "surpassing human performance at 72.4%"; that reference figure comes from the OSWorld paper, which had people attempt its 369 tasks, and a specialised computer-use agent had already passed it the previous December. GPT-5.4 scored 83% on GDPval, which measures realistic knowledge work. The API context window reached one million tokens, the company's longest, and a new "Tool Search" cut the tokens spent on tool calling. OpenAI said individual factual claims were wrong 33% less often than in GPT-5.2, and whole responses 18% less often. It shipped in ChatGPT as GPT-5.4 Thinking, with a GPT-5.4 Pro for more compute on hard problems. On the 17th, smaller GPT-5.4 mini and GPT-5.4 nano followed; mini reached free and Go accounts as well as the API and Codex, while nano was API-only.
9
PolicyAnthropic / US Department of Defense / Donald Trump / Pete Hegseth
Anthropic sues the federal government, calling it unlawful retaliation
Anthropic filed two suits against the Department of Defense and other federal agencies, one in the US District Court for the Northern District of California and one in the US Court of Appeals for the D.C. Circuit. It argued that the late-February "supply chain risk" designation and the government-wide ban on its products went beyond an ordinary contract dispute and amounted to an "unlawful campaign of retaliation" that violated its First Amendment rights and exceeded the reach of the supply chain risk statute. The negotiations had broken down over two conditions Anthropic would not drop: that Claude not be used for mass surveillance of US citizens, and not for killing by autonomous weapons. The company said its "reputation and core First Amendment freedoms are under attack" and asked the courts to stop the bans from being enforced.
11
GovernanceAnthropic / Jack Clark / Sarah Heck
Anthropic sets up the Anthropic Institute under Jack Clark
Anthropic established the Anthropic Institute, a research arm for what it called the most significant challenges powerful AI will pose to societies, led by co-founder Jack Clark in a new role as Head of Public Benefit. It brought together three existing teams — the Frontier Red Team, which pushes models to their limits; Societal Impacts, which studies how AI is actually used; and Economic Research, which tracks effects on jobs and the economy — and added new work forecasting AI progress and examining how powerful systems interact with the legal system. The Institute, Anthropic said, "has access to information that only the builders of frontier AI systems possess." The company also expanded its public policy team, with Sarah Heck taking the role of Head of Public Policy.
17
GovernanceMicrosoft / Satya Nadella / Mustafa Suleyman / Jacob Andreou
Microsoft unifies Copilot, freeing Suleyman for in-house models
Microsoft merged its consumer and commercial Copilot organisations into one, arranged around four pillars: the Copilot experience, the platform, the Microsoft 365 apps, and AI models. Jacob Andreou, who had been corporate vice president for product and growth at Microsoft AI, became executive vice president for Copilot and took charge of the combined team. Mustafa Suleyman stepped away from day-to-day product responsibility to concentrate on superintelligence and Microsoft's own frontier models. Both report to Satya Nadella, who gave as his reason that AI was moving from answering questions and suggesting code to carrying out multi-step tasks, and that a collection of good products had to become one integrated system. It was the largest rearrangement of Microsoft's AI organisation since Suleyman arrived two years earlier.
24
ProductOpenAI
OpenAI announces shutdown of Sora app less than six months after launch
OpenAI announced the shutdown of Sora, its video-generation social app, in a post on X saying "We're saying goodbye to Sora," less than six months after its launch on September 30, 2025. A two-stage timeline followed: the app and web version would close on April 26, 2026, and the Sora API on September 24, 2026, with users urged to download their content before the cutoff dates. Media reported that high inference costs far outweighed revenue while active users declined; OpenAI said it would reallocate compute toward coding and enterprise products, continuing video generation as world-model research.
26
PolicyAnthropic / Rita Lin / US Department of Defense / Emil Michael
A federal judge blocks the Anthropic ban, calling it retaliation
Rita Lin, a US district judge in the Northern District of California, granted a preliminary injunction blocking the "supply chain risk" designation against Anthropic and the ban on federal agencies using its products. The record, she wrote, "strongly suggests that the reasons given for designating Anthropic a supply chain risk were pretextual and that their real motive was unlawful retaliation." What Anthropic had refused to give up were two conditions: that Claude not be used for mass surveillance of US citizens, and not for killing by autonomous weapons. The Pentagon did not comply. Emil Michael, under secretary of defense and chief technology officer, posted on X that the order contained "dozens of factual errors" and that the designation remained "in full force and effect" under the relevant statute and outside the court's jurisdiction. A Pentagon spokesperson declined to say more and pointed reporters to Michael's posts.
26
GovernanceAnthropic / Fortune
A misconfiguration reveals Anthropic's unreleased Mythos
A Fortune reporter found a data store attached to Anthropic's blog infrastructure sitting open to the public — no authentication, fully searchable — holding some 3,000 unpublished assets. Among them was a draft post announcing an unannounced model called Claude Mythos, which said the company regarded it as posing unprecedented cybersecurity risks; the drafts also used the name Capybara. Anthropic confirmed the model existed, calling it "a step change" in performance and "the most capable we've built to date," and said early access customers were trialling it. The cause was a configuration error in its content management system. A company that had built its position on safety had disclosed the existence of the model it considered most dangerous by its own mistake.
31
GovernanceAnthropic / Fortune
Claude Code's source leaks, the second lapse in five days
Claude Code's source code was reported to have been left publicly accessible — some 500,000 lines across 1,900 files — after Anthropic pushed the original source to NPM alongside the built package meant for distribution. Anthropic said "no sensitive customer data or credentials were involved or exposed," calling it "a release packaging issue caused by human error, not a security breach," and said it was putting measures in place to prevent a repeat. A security researcher described it as human error from a shortcut that bypassed the normal release safeguards. It came five days after a misconfiguration on the company's blog infrastructure had exposed the existence of the unreleased Mythos model.

↑ 2026