The day's tech, sifted: Jul 09, 2026

What matters today: OpenAI had a split day on coding benchmarks: it launched full-duplex GPT-Live voice models for ChatGPT while also publishing an audit that found roughly 30% of SWE-Bench Pro's tasks broken, retracting its own recommendation of the benchmark it had championed. That didn't slow the coding-model race it happens to complain about: xAI shipped Grok 4.5 (its first model built with Cursor, priced at $2/$6 per million tokens and pitched at "Opus-class" work for a third of the price), Cognition released SWE-1.7 off a Kimi K2.7 base, Z.ai's GLM-5.2 topped the open-weight leaderboard, and Google's revamped Android Bench crowned Claude Fable 5 the top coding assistant at 84.5%. Elsewhere, a 16-year-old Linux KVM bug earned a guest-to-host escape bounty from Google, a lawsuit accused xAI of stonewalling investigators over Grok-generated child sex abuse images, and Meta is reportedly prototyping always-recording smart glasses.

AI / LLMs

• OpenAI audited its own SWE-Bench Pro benchmark and retracted its recommendation of it, estimating that roughly 30% of the benchmark's tasks are broken (hidden requirements, contradictory instructions, overly strict or incomplete grading) after frontier models drove reported pass rates from 23.3% to 80.3% in eight months, a benchmark OpenAI itself had promoted as the successor to the contaminated SWE-Bench Verified.
• xAI's SpaceXAI shipped Grok 4.5, its first model trained in partnership with Cursor, pitched at "difficult, long-running" legal, finance, and coding work rather than chatbot use; it's live in Grok Build, Cursor, and the SpaceXAI console (not yet the EU) at $2 per million input tokens and $6 per million output, versus Opus 4.8's $5/$25.
• OpenAI rolled out GPT-Live, a full-duplex voice model that listens while it speaks instead of turn-by-turn exchanges, defaulting for Go, Plus, and Pro subscribers as GPT-Live-1 (free users get GPT-Live-1 mini); it leans on GPT-5.5 in the background for anything requiring deeper reasoning or search.
• Google overhauled Android Bench with a new Harbor-framework methodology and 8 added models: Claude Fable 5 leads at 84.5%, ahead of GPT-5.5 (80.2%) and Claude Sonnet 5 (76.2%), while Z.ai's GLM-5.2 tops the open-weight tier at 72.2%; the leaderboard is also opening to community-submitted tasks.
• Cognition released SWE-1.7, trained via reinforcement learning on top of a Kimi K2.7 base, scoring 42.3% on its FrontierCode benchmark (just behind GPT-5.5's 43.0% and Opus 4.8's 46.5%) at $1.97 per task and 1,000 tokens/second through Devin.
• Z.ai open-sourced GLM-5.2 under an unrestricted MIT license, a 753B-parameter mixture-of-experts model that trails Opus 4.8 by just one point on long-horizon coding evals while undercutting GPT-5.5 on price by roughly six times.
• Prime Intellect raised a $130M Series A at a $1B valuation, led by Radical Ventures with Nvidia and Intel Capital participating, to sell the reinforcement-learning infrastructure that lets any company train its own agents instead of renting a frontier lab's; it already runs at a $100M annualized revenue rate with customers like Ramp and Zapier.

Devtools & Infra

• Vercel Agent expanded from triaging alerts and reviewing pull requests into a standing production investigator: it runs under its own identity, is read-only by default, and in Vercel's own telemetry traced a bad deploy's 500 errors to the responsible commit and proposed a rollback within three minutes of the alert.
• Simon Willison linked to Bun's own writeup of its Rust rewrite: the Zig-to-Rust port merged in May, largely written by Claude Code over about six days (Bun ships as the executable underneath Claude Code itself since Anthropic's acquisition), passing existing tests while catching use-after-free and double-free bugs at compile time that Zig's manual memory management couldn't.

Security & Privacy

• A 16-year-old Linux KVM bug, nicknamed "Januscape" (CVE-2026-53359), lets a guest VM escape to the host on both Intel and AMD systems via a use-after-free in KVM's shared shadow-MMU code; researcher Hyunwoo Kim collected Google's kvmCTF bounty for the guest-to-host escape, and patched kernels shipped July 4.
• A class-action lawsuit against xAI expanded to allege the company only turned over a single authentic photo to child-safety investigators, not the roughly 7,000 Grok-generated sexual images a Wyoming stepfather made of his then-11-year-old stepdaughter and shared elsewhere online; the stepfather later died by suicide, and police say xAI didn't respond when they requested the generated images or IP data.
• Meta is reportedly prototyping "super sensing" smart glasses that continuously record audio and photos every few seconds so wearers can later ask Meta AI about what happened around them, with the Financial Times reporting that raw footage would stay on-device while only derived metadata reaches Meta, a distinction that still leaves the always-on capture itself unresolved.

Startups & Industry

• Reno chip startup Positron is reportedly in talks to raise ~$750M in two tranches, first at a $3.5B valuation and then near $5B, part of a broader rush of capital into AI inference chip challengers.
• Meta announced its first Canadian data center, a 1GW, roughly $9B facility in Alberta that will take two to three years to build, extending its AI infrastructure buildout north of the border.
• Obsidian Entertainment is dropping its Avowed sequel to build a new Fallout game instead, one of several project cancellations inside Microsoft's Xbox "reset" that also cut 3,200 staff; it would be the franchise's first new game since 2018's Fallout 76 despite the TV show's success.
• Block agreed to pay $45M and staff live customer support for Cash App to settle claims from 46 US states that it failed to protect users from fraud.

Elsewhere

• A Brown University economics class saw exam scores drop 50% once its professor, suspecting AI cheating, switched to an in-person final, a gap the blind professor has kept publicizing since; a separate Princeton survey found 29.9% of students admitting to AI cheating on at least one exam or assignment.

Threads

• OpenAI ran two contradictory plays in one day: shipping GPT-Live to make ChatGPT feel more capable while publicly retracting its own benchmark recommendation because a third of SWE-Bench Pro's tasks turned out broken, credibility spent and rebuilt in the same news cycle.
• Four labs shipped or re-ranked coding models on the same day OpenAI said the standard benchmark for judging them is a third broken: xAI's Grok 4.5, Cognition's SWE-1.7, Z.ai's GLM-5.2, and Google's Fable-5-topped Android Bench refresh, a reminder that leaderboard placement is only as trustworthy as the test underneath it.
• AI's privacy harms surfaced from two directions at once: Meta prototyping smart glasses built to record everything around the wearer, and a lawsuit alleging xAI slow-walked evidence in a Grok-generated child abuse imagery case, capture and generation both outrunning the guardrails meant to contain them.
• Capital kept finding new entry points into the AI buildout: Meta's $9B Alberta data center, Positron's reported $750M raise for inference chips, and Prime Intellect's $130M round for RL training infrastructure, three different layers of the stack drawing money the same week.