GPT-5.4 Hits 83% Human-Level Performance as Amazon Data Center Becomes Act-of-War Target

OpenAI released GPT-5.4 Thinking and GPT-5.4 Pro with native computer use, a million tokens of context, and an 83.0% GDPval score, matching or exceeding human professionals at knowledge work 83% of the time.

Other benchmark results: 57.7% on SWE-Bench Pro (software engineering), 67.3% on WebArena-Verified (web tasks), and 54.6% on Toolathlon (tool use).

The model built a "basically perfect" Minecraft clone in 24 minutes, apparently the first AI to do so.

On FrontierMath Tier 4, the hardest category of research-level math problems, GPT-5.4 Pro posted 38.0%. Mathematician Bartosz Naskrecki, whose Tier 4 problem was solved by AI for the first time, called it his personal "move 37," referring to the famous unexpected move AlphaGo made against Lee Sedol. He said the solution was clean and felt almost human.

Epoch AI reported the model has begun making novel observations on FrontierMath's Open Problems benchmark, where no model has scored before.

Netflix has acquired InterPositive, Ben Affleck's AI filmmaking startup that trains models on production dailies to handle mixing, coloring, relighting, and VFX in post.

Apple Music is launching "Transparency Tags" that require labels specifying whether a track's artwork, composition, or music video involved AI generation.

Apple has quietly pulled the 512-GB RAM Mac Studio configuration as the DRAM shortage reaches Cupertino.

Washington is drafting regulations requiring American approval for AI chip shipments anywhere on Earth. Commerce proposes that nations wanting Nvidia and AMD chips invest in America as a condition of access.

Cloverleaf Infrastructure, a Seattle startup selling "powered land" for data centers, raised $300 million.

Oracle is cutting thousands of jobs to fund its data center expansion, targeting roles it expects AI to replace.

TerraPower received the first-ever NRC construction permit for a commercial-scale advanced nuclear plant, a 345-MW sodium-cooled fast reactor in Wyoming with molten salt energy storage. A sodium-cooled fast reactor uses liquid sodium instead of water as a coolant, allowing it to operate at higher temperatures and potentially use nuclear waste as fuel.

Iranian state media revealed that Amazon's Bahrain data center was deliberately targeted by the IRGC for the company's US military support, making it one of the first known cases of a data center intentionally struck as an act of war.

A University of Chicago economist finds that AI's impact on productivity is now visible in macro aggregate data.

Citadel Securities published a chart showing software engineer job postings surging, a Jevons paradox where efficiency breeds more demand. The Jevons paradox is when improvements in efficiency lead to increased total consumption rather than decreased consumption.

Anthropic launched "observed exposure," an early-warning system for AI-driven white-collar job displacement that weights automated over augmentative uses. So far it suggests limited evidence of job loss.

New York is considering a bill to ban chatbots from giving legal and medical advice, with a private right of action for violations.

Indiana's governor signed a law requiring cryptocurrency options in investment and retirement plans by July 2027.

The Pentagon has formally designated Anthropic a supply-chain risk, a classification likely to force military contractors to sever ties.

Dario Amodei responded that Anthropic has been having "productive conversations" with the DOD, and apologized for a leaked memo that criticized the White House and Sam Altman.

A suspected military insider has won $90,000 on Polymarket by correctly predicting US strikes on Venezuela, Iran, and Iraq, and now bets that American forces will enter Iran this year.

NASA's Nancy Grace Roman Space Telescope, with a field of view 100 times larger than Hubble's, is set to launch this fall.

Vast has raised $500 million to build space stations from low-Earth orbit to the Moon and Mars.