The day's tech, sifted: Sep 6, 2026

What matters today: Artificial Analysis's rebuilt Intelligence Index v4.2 returned its first scores: Claude Fable 5.1 leads GPT-6 Astra 57 to 55, the first independently-scored number to land since yesterday, when Simon Willison flagged Astra's 99.9% ARC-AGI-3 claim as harness-dependent. That skepticism gained more weight today: OpenAI has quietly revised several of Astra's own evaluation benchmarks since the model's Sept 3 launch, changes Fortune reports appear to favor Astra. The same scrutiny caught LlamaIndex's ExtractBench finding GPT-5 and Gemini-3 both score 0% on its widest document-extraction schemas.

AI / LLMs

• Artificial Analysis's rebuilt Intelligence Index v4.2 returned its first scores: Claude Fable 5.1 leads at 57, GPT-6 Astra second at 55 (a 4-point gain over GPT-5.6 Sol), the first hard number to arrive since Astra's contested 99.9% ARC-AGI-3 score.
• OpenAI has quietly revised several of Astra's evaluation benchmarks since first publishing its Sept 3 announcement, per Fortune, changes that appear to favor the model and follow yesterday's criticism that its ARC-AGI-3 score was harness-dependent.
• Zvi's roundup of the community verdict on Fable 5.1 and Mythos 5.1 finds near-universal praise for writing quality and fewer intrusive safety refusals (Anthropic says biology-safeguard interventions on benign requests dropped 85%), alongside a 75% cut to cache-read pricing ($1 to $0.25 per million tokens) that Claude Code creator Boris Cherny says cuts a typical Claude Code session up to 38%.
• LlamaIndex and Kaggle's new ExtractBench grades document extraction across 370 enterprise files on missing fields, source grounding and duplicate records, and found GPT-5 and Gemini-3 both score 0% on its widest schemas, a sign leaderboard-topping models still fail basic structured-output work.
• Meta's autonomous research agent AIRA3 placed 8th of roughly 4,000 teams in an Nvidia-run Kaggle contest to fine-tune Nemotron, gold-medal territory for an agent competing unassisted against human teams.
• MBZUAI released K2 Horizon, six fully open-source models from 0.9B to 375B parameters, complete with weights, code, training data and recipes under Apache 2.0, one of the largest fully open model releases to date.

Startups & Industry

• Swiss Re expects global data-center insurance premiums to reach $20B-$30B a year by 2030, with roughly 40% of US data-center capacity sitting in tornado-prone areas, per the Wall Street Journal, one more sign AI buildout risk is migrating onto insurers' books.
• The data-center backlash is testing Texas's famously pro-business stance, even as Wood Mackenzie finds the state has more data-center capacity under construction than any other, per the Financial Times.
• The Seattle Times and Newsday sued OpenAI and Microsoft, alleging the companies trained their models on the papers' journalism without permission; Microsoft and OpenAI are also funders of the Seattle Times, an awkward detail in an already crowded field of publisher AI lawsuits.
• Travis Kalanick's robotics startup Atoms is developing robotaxi tech and has hired Anthony Levandowski after acquiring his self-driving startup Pronto, per the Financial Times; Uber, the company that ousted Kalanick a decade ago, has invested $100M in Atoms.

Research

• As researchers apply AI to decode animal communication, bioethicists warn the same tools could give humans new ways to manipulate or exploit other species, per Bloomberg, a caution note on a research area usually framed as pure upside.

Elsewhere

• iOS 27's new "iPhone Handoff" feature will let two iPhones share one phone number, switching an active line between a primary device and a "companion" phone, timed to land alongside Apple's expected foldable iPhone Ultra.
• The Trump administration launched arcade.gov with a deportation-themed Snake clone and a Tetris-style "Build the Wall" game where players fend off zombies at the border; Tetris's rights holder says it takes the IP infringement "very seriously."
• Earth Garden turns a spinning 3D globe into a browser for field recordings, from island surf to Tanzanian cricket choruses, a Radio Garden for ambient sound instead of radio stations.
• Isar Aerospace's privately-developed Spectrum rocket became the first fully commercial launch vehicle in Europe to reach orbit, launching Saturday from a Norwegian spaceport inside the Arctic Circle and successfully deploying payloads on its second flight.
• AI has gutted Kenya's essay-writing gig industry, which at its peak paid more than 40,000 people in Nairobi to write overseas students' homework, per the New York Times, leaving few paths back to work for a labor market AI ate almost overnight.

Hacker News

Isar Aerospace's orbital launch (covered above) also topped Hacker News's board today. Security folks are passing around a report that qBittorrent broke out of its sandbox to run arbitrary commands, notable given the client's install base. On the self-hosting side, Cloud in a Bottle launched aiming to make running your own infrastructure approachable, and a packagemain post arguing .gitignore should default to ignoring everything and allowlist tracked files instead drew heavy engagement. Chrome watchers flagged that the browser still carves out exceptions for Google's own site data under settings meant to apply uniformly.

AI fatigue ran through several threads: LLMs as a Cognitive Virus frames chatbot use in epidemiological terms, the hobbyist community A/I announced it's shutting down, and a personal essay on how AI makes the author feel got unusually broad traction alongside news of OpenAI's GPT-6 Astra running on robot arms. Separately, Wikimedia Foundation staff voted overwhelmingly to unionize with CWA, the latest tech-labor org drive to clear a supermajority.

Threads

• Benchmark claims kept needing hard evidence instead of vendor pitches: yesterday Simon Willison flagged Astra's 99.9% ARC-AGI-3 score as harness-dependent, today OpenAI's own quiet revisions to Astra's eval benchmarks lent that skepticism more weight, Artificial Analysis's independently-scored index put daylight between Fable 5.1 and Astra with a checkable number, and ExtractBench caught both companies' flagships failing completely on a task neither had bragged about.
• The data-center buildout is becoming a two-front problem: insurers are pricing in tornado risk at the same time local backlash is testing Texas's pro-business welcome for the industry.
• Government and tech collided twice in unrelated ways: two newspapers sued OpenAI and Microsoft over training data, while Tetris's rights holder threatened the federal government itself over a cloned arcade game on arcade.gov.
• AI is displacing human work at both ends of the skill ladder: Meta's AIRA3 agent placed 8th of roughly 4,000 human teams in an elite Kaggle research contest, while AI has gutted the essay-writing gig economy that once paid 40,000+ Kenyans to do overseas students' homework.