{"issue":{"id":"2026-09-21","coversLabel":"August 1 – September 20, 2026","number":1,"readMinutes":4,"weekOf":"2026-09-21","publishedLabel":"Week of September 21, 2026","title":"AI agents get cheaper as new disclosure rules land","dek":"Anthropic and OpenAI shipped models built to work on their own, Gemini learned to talk and act at once, EU AI disclosure duties took effect, and California sent an AI firing bill to the governor.","companyLabel":"Public edition","models":[{"slug":"anthropic-claude-fable-5-1","kind":"model","source":{"title":"Introducing Claude Fable 5.1 and Claude Mythos 5.1","url":"https://www.anthropic.com/claude-fable-and-mythos-5-1","date":"2026-09-01"},"headline":"Cheaper cache reads cut the bill for long-running Claude agents by up to 45%","whatChanged":"Anthropic released Claude Fable 5.1 on September 1. It costs $10 per million input tokens and $50 per million output tokens, and the price of cache reads fell 75%, to $0.25 per million. Anthropic estimates typical workloads get about 25% cheaper and agent-style work up to 45% cheaper. A second model, Claude Mythos 5.1, is limited to vetted U.S. organizations. Enterprise Frontier Safeguards, which pair zero data retention with cloud infrastructure the customer controls, roll out in phases starting this fall.","whyItMatters":"Agents reread the same files over and over, so the cache discount is where most of the savings land. That makes it realistic to leave an agent working on a contract archive, a support backlog, or a vendor review for days. The harder question is oversight. Someone who starts an agent on Friday and reads its output on Monday has handed off everything in between. Teams should agree on which tasks can run unattended and who checks the result before anyone relies on it.","issuer":"Anthropic","glance":"Anthropic released Claude Fable 5.1 on September 1 and cut cache-read prices by 75%, to $0.25 per million tokens.","keyFact":"Cache reads down 75%, to $0.25 per million","action":"Pick the tasks your team may hand to an unattended agent, and name who checks the output.","companyLabel":"Public edition"},{"slug":"openai-gpt-6-astra","kind":"model","source":{"title":"GPT-6 Astra: A new generation of intelligence","url":"https://openai.com/index/gpt-6-astra/","date":"2026-09-03"},"headline":"GPT-6 Astra can operate a browser, and enterprise admins decide who gets it","whatChanged":"OpenAI began rolling out GPT-6 Astra on September 3. OpenAI calls it state of the art at computer use, browsing, software engineering, cybersecurity, and professional work. On enterprise and API accounts it stays off until an administrator enables it. The API model, `gpt-6-astra`, costs $10 per million input tokens and $50 per million output tokens, with cached input at $1 and batch jobs at half price. It is also available in ChatGPT plans.","whyItMatters":"A model that uses a computer can go anywhere its user's login goes: email, the CRM, the finance system, a customer portal. The risk shifts from what the model writes to what it does. Because Astra starts switched off for business accounts, there is a clean moment to decide before anyone uses it. That call sits with whoever holds admin rights, and it works best as a deliberate choice about which systems an agent may touch and how its actions get logged.","issuer":"OpenAI","glance":"OpenAI began rolling out GPT-6 Astra on September 3, a flagship built for computer use and off by default for enterprise accounts.","keyFact":"Off until an administrator turns it on","action":"Find out who holds admin rights in your OpenAI workspace and agree who approves turning Astra on.","companyLabel":"Public edition"},{"slug":"google-gemini-3-8-live","kind":"model","source":{"title":"Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/","date":"2026-09-15"},"headline":"Personal Google AI plans now put a tool-using voice assistant inside Gmail and Docs","whatChanged":"Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15. Both hold spoken conversations in 97 languages while calling tools in the background, and Extended Thinking reasons as it speaks. Google AI subscribers can use Extended Thinking in Gmail and Keep, and Pro and Ultra subscribers can use it in Docs. Search Live runs on 3.8 Live. Businesses get a private preview in Gemini Enterprise, and generated audio carries a SynthID watermark.","whyItMatters":"The consumer rollout comes first, and that is where company data can slip out. An employee paying for Google AI on a personal account can now talk through a work problem with an assistant sitting inside their own email and documents. Spoken requests that end in a sent email or an edited file also leave a thinner record than typed ones. Business use is still in private preview, which gives IT time to set the rules for work accounts and for personal ones.","issuer":"Google","glance":"Google released Gemini 3.8 Live on September 15, a voice model that keeps talking while it runs tools, now in Gmail, Docs, and Search.","keyFact":"In Gmail, Docs, Keep, and Search Live","action":"Check which Gemini features are switched on for work Google accounts, and write down the rule for personal ones.","companyLabel":"Public edition"}],"rules":[{"slug":"eu-ai-act-article-50-transparency","kind":"rule","source":{"title":"Guidelines on transparency obligations for providers and deployers of AI systems","url":"https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems","date":"2026-07-20"},"headline":"Chatbots and synthetic media reaching the EU now carry disclosure duties","whatChanged":"Article 50 of the EU AI Act has applied since August 2. Providers must make it clear when a person is interacting with an AI system, unless that is obvious, and must mark certain AI-generated audio, images, video, and text so machines can detect them. Deployers have their own duties, including disclosing deepfakes and some AI-generated text. The European Commission published guidelines on July 20 on how it expects the duties to be met. The guidelines interpret the law; the duties themselves are binding.","whyItMatters":"Article 50 can reach a U.S. company when its AI system's output is used in the EU. That covers customer-service chat, voice assistants, avatars, and AI-drafted content that EU residents see. A company in scope needs disclosure wording and a way to mark synthetic media, and it helps to keep a record of what each customer was shown. Whether a specific tool is covered is a question for counsel, and the inventory can start now.","issuer":"European Union","glance":"EU AI Act Article 50 has applied since August 2: people must be told when they are dealing with an AI system.","status":"in-force","action":"List customer-facing AI tools that EU residents could use, and check that each one says it is AI.","companyLabel":"Public edition"},{"slug":"california-sb-947-automated-decisions-employment","kind":"rule","source":{"title":"SB-947 Employment: automated decision systems","url":"https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202520260SB947","date":"2026-09-04"},"headline":"Using AI to fire California workers could soon require a human sign-off","whatChanged":"California's legislature passed SB 947 and enrolled it on September 4. The governor has until September 30 to sign or veto it. If signed, from July 1, 2027, employers could not rely primarily on an automated decision system to discipline or fire a worker unless a human corroborates the decision. They also could not use such a system to infer protected status or punish workers for exercising legal rights. Affected workers would get a written notice afterward, and the Labor Commissioner could issue $500 citations per violation.","whyItMatters":"Plenty of companies now feed productivity scores, scheduling data, or monitoring alerts into performance decisions. If SB 947 is signed, discipline or dismissal of a California employee that rests mainly on one of those tools would need a person to corroborate it and a notice to the worker. HR, managers, and whoever owns those tools would need a shared review step before mid-2027. The bill is not law yet, and the governor's decision sets the clock.","issuer":"California Legislature","glance":"California sent SB 947 to the governor on September 4; it would require human review of AI-driven firing and discipline.","status":"proposed","action":"List which discipline decisions for California staff start from an automated score or alert.","companyLabel":"Public edition"}],"older":null,"newer":{"id":"2026-09-28","publishedLabel":"Week of September 28, 2026","title":"Cheaper top-tier AI meets a warning about agent liability","isLatest":false}},"session":null}