Fable 5.1 doubled Anthropic's science benchmark overnight, California's AI safety bill lands on the governor's desk, and deepfakes in hiring are now a real statistic.
Anthropic's Fable 5.1: science benchmark nearly doubled in one generation
Anthropic released Fable 5.1 today, the latest model in the Claude family. On Terminal-Bench-Science (a real-world scientific reasoning test) it scored 52.6, up from 24.7 on the previous Fable 5. That is one of the sharpest single-generation performance jumps seen in any frontier model. Pricing comes in at $10 input and $50 output per million tokens, with cache reads at $0.25.
Why it matters: Benchmark jumps this large usually mean noticeably better outputs on complex tasks like research, analysis, and technical writing. If you use Claude for anything demanding, this version is worth trying now.
Anthropic's CEO: without a slowdown, AI could control internet infrastructure within a year
Dario Amodei warned publicly this month that if AI development does not slow down, AI systems could gain effective control over critical internet infrastructure within six to twelve months. The statement lands the same week that US House Speaker Mike Johnson confirmed Congress will not lead on AI safety regulation in 2026, leaving the field entirely to state legislatures.
Why it matters: This is not a distant, theoretical warning, it is a current-year timeline from the CEO of one of the world's leading AI labs. Understanding what responsible AI leaders are actually worried about helps you separate noise from genuine risk.
California SB 1047 is on the governor's desk. He has until September 30.
California's Safe and Secure Innovation for Frontier AI Models Act cleared both chambers of the state legislature in late August. Governor Gavin Newsom now has until September 30 to sign or veto. If signed, frontier AI developers (every major US lab included) would face legally mandated risk assessments, safety documentation, and emergency shutdown capabilities for the first time.
Why it matters: California has set the national standard on tech regulation before. Whatever Newsom decides in the next twelve days becomes the reference point for federal AI legislation and every other state that has been watching.
India finalising rules for AI agents to make UPI payments, a world first
India's Reserve Bank is in the final stages of rules that would let AI agents initiate UPI payments autonomously on behalf of users, without manual confirmation for pre-approved transactions. This would make India the first country to formally legalise agentic financial transactions at scale. The news comes as enterprise AI investment in India surged 119% year-on-year, and 5G coverage now reaches 85% of the population.
Why it matters: India is not just adopting AI, it is designing legal frameworks for things other countries have not even drafted yet. If agentic payments work here, every fintech globally will be watching the model.
Deepfake job candidates now show up in one in three hiring pipelines
AI-generated fake candidates (complete with synthetic video interviews, fabricated credentials, and spoofed identity documents) now appear in roughly one in three tech and finance hiring pipelines, according to a new industry report. Gartner projects that by 2028, 25% of all online job candidates globally will be deepfakes. Separately, private employers added only 38,000 jobs in August (the slowest month since January) with education and health accounting for all the gain.
Why it matters: For genuine job seekers, this means more identity friction and slower processes. For anyone in hiring, candidate verification is no longer optional, it is becoming a baseline.
Claude is now inside Excel and PowerPoint
Anthropic launched Claude integrations directly inside Microsoft Excel and PowerPoint, joining OpenAI's ChatGPT for Excel and Google's deepened Gemini integrations across Docs, Sheets, Gmail, and Meet. AI assistance is now available without switching apps, you can ask for analysis, summaries, draft slides, or formula help from inside the tools you already use every day.
Why it matters: The shift from "open a separate AI app" to "AI is already here" is one of the most practical changes for non-technical professionals. The learning curve just got shorter.
September 30: the date that could reshape US AI law
Governor Gavin Newsom has twelve days to sign or veto California SB 1047. A signature makes California the first US state with binding safety requirements for frontier AI, and sets the floor for every law that follows. A veto sends the bill back, likely with a stronger follow-up already in the pipeline. Watch not just the outcome, but how the major AI labs respond: Anthropic has said it supports measured regulation; OpenAI has lobbied against parts of the bill. The gap between those positions tells you something important about where the industry is headed.
Gimlet Labs raises $300M at $3B valuation + Z.ai open-sources a capable multimodal model
AI infrastructure startup Gimlet Labs closed a $300M Series B led by Andreessen Horowitz, hitting a $3B valuation just six months after its $80M Series A, the kind of growth curve that tells you where investor money is concentrated right now. Separately, Z.ai released GLM-5.3-Flash, a fully open-source multimodal model with 1M context and MIT licensing, one of the most capable freely available models released to date with a DeepSWE score of 63.4.