Daily AI brief · September 6, 2026
Benchmark Overhaul Puts Fable 5.1 Ahead of GPT-6 Astra
Artificial Analysis rebuilt its Intelligence Index with more private tests and a long-PDF benchmark, ranking Claude Fable 5.1 first and GPT-6 Astra second, while courts and lenders shaped the rest of the week.
Artificial Analysis overhauls its Intelligence Index; Fable 5.1 leads, Astra second (Trending Topics, The Neuron) Version 4.2 of the widely watched index doubles the weight of private held-out tests to 40%, drops the saturated GPQA Diamond, and adds two new benchmarks: AA-Briefcase for agentic knowledge work and GDP.pdf, which grades reasoning across 100 real PDFs totaling 4,592 pages. Claude Fable 5.1 took first place and GPT-6 Astra second, but Astra led the PDF test with a 33.2% pass rate to Fable 5.1's 26.2%. What it means for SMBs / MSPs: The overall ranking matters less than the sub-scores; if your clients' work is document-heavy, the PDF result is the number to watch.
xAI loses its bid to block Minnesota's AI "nudification" ban (Reuters via U.S. News, The Star) U.S. District Judge Donovan Frank denied xAI a preliminary injunction against the law, which took effect August 1 and bars operators and developers from letting users "nudify" images of identifiable people with AI. The judge found xAI hadn't shown it would be harmed while its First Amendment suit proceeds; xAI said it will appeal to the 8th Circuit. What it means for SMBs / MSPs: State AI laws aimed at specific harms are surviving early challenges, so image tools your clients embed in apps or sites need clear use restrictions now.
ByteDance signs a $29.6 billion loan to fund AI outside China (Quartz, Tech Startups) TikTok's parent signed an unsecured three-year facility with 28 banks, upsized from a $20 billion target, to fund AI work beyond China's borders, including data centers under development in Southeast Asia. It is reported as Asia's second-largest loan of the year. What it means for SMBs / MSPs: Capacity keeps piling up worldwide, which supports the case that AI prices will keep falling even as flagship models get more expensive per token.
Your own GDP.pdf test. The new index shows models differ most on long, messy documents, and that's exactly what small businesses feed them: contracts, vendor quotes, insurance policies. This is for the champion choosing a model for document work, and a first run takes about 45 minutes. First step: pick three real PDFs of 20 pages or more, write five questions per document with known answers, and run them through the two or three models you already have access to. Score correct, partly correct or wrong, and keep the sheet for the next model release.
Add one rule for your team: a leaderboard change is not a reason to switch AI tools. The Artificial Analysis index was just rebuilt, so rankings moved partly because the test changed. Its sub-scores also disagree: Fable 5.1 ranked first overall, but GPT-6 Astra led the PDF test 33.2% to 26.2%. If your work is contracts, quotes and policies, judge tools on your own documents.
*For MSPs:* A fixed-fee document benchmark on ten of a client's PDFs turns model debates into a one-page scorecard that can be rerun when a new model ships.
Switching models every time a leaderboard reshuffles. This index was just rebuilt, so rankings moved partly because the test changed, not because the models did. Change your default model when your own tests say so, not when a chart does.
Benchmarks are moving toward real work, and behind closed doors. Private held-out tests, long real-world PDFs and agentic task suites are replacing public quiz-style exams that models have memorized or saturated. That makes leaderboards more useful, but also harder to check. The practical result for smaller buyers: public scores will point you to a shortlist, and your own small, repeatable test set will make the final call.
Compiled on September 27, 2026 from reporting published on or before September 6, 2026.