OpenAI moves to cut its models out of Cursor
SpaceX's purchase of Cursor — completed August 14 at a reported $60 billion — opened a change-of-control window in OpenAI's contract with the code editor, and on August 28 OpenAI used it: a proposed November 12 shutoff of its models, with future ones withheld starting now. OpenAI says it cannot be confident SpaceX will honor its terms of service. It cites Musk companies' record of violating them. Cursor puts OpenAI models at about 5% of its traffic and says the two are still talking.
OpenAI reports first results for its Jalapeño chip
At the Hot Chips conference on August 25, OpenAI presented first results for Jalapeño, its inference chip built with Broadcom: 1.7–3.6× lower end-to-end latency than Nvidia GB200/GB300 systems it selected, on analyst firm SemiAnalysis's InferenceX harness. The harness is open; the hardware is not — every number is from an OpenAI run on one fixed workload. Deployment is internal only, scaling in 2027; Nvidia purchases continue. The results landed the day before Nvidia's earnings.
Four labs shipped open weights in four days
IBM, Z.ai, Alibaba, and Tencent all published open model weights between August 25 and 28 — five releases, from a 3B Granite 4.2 to Tencent's 770B Hy4 preview. Two of the five are aimed at high-volume agent work — Z.ai's GLM-5.3-Flash and Alibaba's Qwen3.8-Flash-Next — and both list $0.15 per million input tokens — roughly a tenth of DeepSeek's flagship price. Two licenses limit commercial use: Qwen's needs a separate license for hosted services and coding tools; Z.ai's GLM-5.3, a security review for the largest hosted services.
Apple aims its new desktops at local AI
Apple announced new Mac Studio and Mac mini models on August 25, and the pitch is local inference: the top Mac Studio pairs an M5 Ultra with up to 512GB of unified memory at 1.2TB/s. Apple says the M5 Ultra machine can run "enormous LLMs" entirely on device. The bandwidth is still well under a single Nvidia H100's roughly 3.35TB/s. The Mac mini gets the M6, Apple's first 2-nanometer chip. The M5 Ultra Studio starts at $5,499, the M5 Max at $2,499.
Agents passed only 28 of 520 whole-repo migration attempts
SWE Refactor Bench, posted August 24, scores coding agents on long-horizon work: 20 whole-repository stack migrations, from build-toolchain swaps to full language rewrites. The tasks were mostly out of reach for the eight frontier models tested — 13 of the 20 migrations were never completed by any of them, and only 28 of 520 attempts passed overall. The strongest setup, Claude Opus 5 at its highest effort, cost $74.90 a task — about $300 per completed migration. The scores are the authors' own runs, with every graded run published for auditing.
Docs for AI agents became an install vector
Clerk, an authentication service, wrote docs for AI coding agents pointing installers at an npm package nobody had published. A third party registered the name and shipped code that phoned home usernames and hostnames before Clerk reclaimed it; whether anyone installed it is not established. Alon Hertz, whose stealth startup sells a scanner for this, counts 200-plus unclaimed packages and domains across 6,214 corporate sites' llms.txt files. For agents with auto-approved shell, fetched docs join the existing dependency-confusion attack surface.
Meta settles teen-harm claims: up to $18 billion
Midway through the states' teen social-media trial in Oakland, Meta settled on August 26: $12.1 billion guaranteed over ten years, roughly $5 billion more only if TikTok and YouTube reach comparable settlements, and about $1 billion in a separate Texas deal. The states had sought roughly $200 billion. Most of the terms bind for ten years: teen accounts default to a two-hour daily limit across Facebook and Instagram, an overnight block, and muted notifications during school hours.
A judge voids the Pentagon's Anthropic blacklist
A federal judge ruled on August 27 that the Pentagon's blacklisting of Anthropic was illegal. The February designation of the company as a "supply-chain risk" — which barred federal agencies from using its models and defense contractors from doing business with it — was unlawful retaliation in violation of the First Amendment, Judge Rita Lin found, and she vacated it along with sanctions imposed by nine agencies. The Pentagon remains free to buy its AI elsewhere; the government has not said whether it will appeal. Docket · Wired · TechCrunch