{"issue":{"id":"2026-09-28","coversLabel":"September 21–27, 2026","number":2,"readMinutes":4,"weekOf":"2026-09-28","publishedLabel":"Week of September 28, 2026","title":"Cheaper top-tier AI meets a warning about agent liability","dek":"Anthropic and OpenAI cut prices, Google gave voice agents a face, the FTC chair said companies answer for their agents, and a Senate bill would make AI labs publish model cards.","companyLabel":"Public edition","models":[{"slug":"anthropic-claude-opus-5-5","kind":"model","source":{"title":"Claude Opus 5.5","url":"https://www.anthropic.com/claude-opus-5-5","date":"2026-09-22"},"headline":"Anthropic says Opus 5.5 matches Fable 5.1 on most work for 40% less than Opus 5","whatChanged":"Anthropic released Claude Opus 5.5 on September 22. Anthropic says it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. List prices are $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million. It is available on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure as `claude-opus-5-5`, and it can be used with zero data retention.","whyItMatters":"Cost has been the main reason teams kept agents on short leashes. At these prices, an agent that works through a full contract set, a support queue, or a quarter of invoices overnight fits an ordinary budget, so expect requests to use it. The data questions stay the same at any price. Before company or customer information goes in, someone should know where prompts and outputs are stored and whether the account actually has zero data retention switched on.","issuer":"Anthropic","glance":"Anthropic released Claude Opus 5.5 on September 22, rated at Fable 5.1's level for 40% less than Opus 5.","keyFact":"$4 in, $20 out per million tokens","action":"Check whether your Anthropic account has zero data retention before approving Opus 5.5 for real data.","companyLabel":"Public edition"},{"slug":"openai-gpt-6-sol-luna","kind":"model","source":{"title":"OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more","url":"https://venturebeat.com/technology/openai-releases-gpt-6-sol-and-luna-models-slashing-api-costs-50-or-more","date":"2026-09-22"},"headline":"OpenAI cuts API prices by half or more with GPT-6 Sol and Luna","whatChanged":"OpenAI released GPT-6 Sol and GPT-6 Luna on September 22. Sol costs $2 per million input tokens and $10 per million output tokens. Luna, the fastest and cheapest, costs $0.10 and $0.50, about half the price of the GPT-5.6 Luna it replaces, with a 90% discount on repeated input. Both are in the API and rolling out across ChatGPT Business, Enterprise, and individual paid plans. OpenAI reported that Sol's deception rate on its internal adversarial tests fell to 1.3%, from 10.4% for GPT-5.6 Sol.","whyItMatters":"Two things follow. Cheap models spread fastest through personal subscriptions, so employees now have a capable GPT-6 model on the same phone they use for work email, outside anything the company has reviewed. And at Luna's price, teams will add AI to every call summary, ticket, and alert. The lower deception rate is good news, and it comes from OpenAI's own tests. Each new model should go through the company's own review before it touches customer or company records.","issuer":"OpenAI","glance":"OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, cutting API prices by half or more.","keyFact":"Sol: $2 in, $10 out per million tokens","action":"Remind staff that work data goes only into company-approved AI accounts, and add Sol and Luna to your AI list.","companyLabel":"Public edition"},{"slug":"google-gemini-3-8-live-avatar","kind":"model","source":{"title":"Gemini 3.8 Live with Live Avatar","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-with-live-avatar/","date":"2026-09-24"},"headline":"Google's Live Avatar lets businesses put a talking, lip-synced face on an AI agent","whatChanged":"Google added Live Avatar to Gemini 3.8 Live on September 24. It generates a talking video avatar in near real time, with lip-sync and facial expression across 97 languages, while the agent behind it keeps calling tools in the background. It launched in Gemini Enterprise, with API access for developers. Organizations can build a custom avatar from a reference image, currently only through enterprise allowlisting. Google says all audio and video output carries a SynthID watermark.","whyItMatters":"There are two sides to this. A company that puts an avatar in front of customers is speaking through it, so what it says needs the same review as any other customer communication. On the other side, a convincing face on a video call is weaker proof of identity than it used to be. SynthID only marks Google's own output, and staff cannot see it anyway, so payment changes and new payees need a check through a channel set up before the request arrived.","issuer":"Google","glance":"Google launched Live Avatar on September 24: real-time talking video avatars in Gemini Enterprise, watermarked with SynthID.","keyFact":"Lip-synced video avatars in 97 languages","action":"Require a call-back to a known number before acting on payment or account changes requested by video.","companyLabel":"Public edition"}],"rules":[{"slug":"ftc-chair-ai-agent-liability","kind":"rule","source":{"title":"FTC Chair Pushes Back on Treating AI Agents as Independent Actors","url":"https://money.usnews.com/investing/news/articles/2026-09-25/reuters-next-ftc-chair-pushes-back-on-treating-ai-agents-as-independent-actors","date":"2026-09-25"},"headline":"If your AI agent misbehaves, the FTC chair says the blame is yours","whatChanged":"FTC Chairman Andrew Ferguson spoke with Reuters at a conference in Austin on September 25. He said he will resist treating AI agents as independent actors with wills of their own. Audit trails of supposedly rogue agents, he said, keep showing them carrying out the instructions they were given, so responsibility sits with the people who instruct them. He added that new laws should wait until existing ones prove insufficient. These were interview remarks, not a policy statement or rule.","whyItMatters":"Any company letting agents send customer emails, issue refunds, or change records can read this as a preview of how the FTC would approach a problem: it would look at the company, under consumer-protection law that already exists. The audit-trail point is the practical one. If a team cannot reconstruct what an agent was told and what it did, it has no way to show a mistake was a one-off. The remarks carry no legal force on their own.","issuer":"FTC","glance":"FTC Chair Andrew Ferguson said on September 25 that the people who instruct AI agents are responsible for what the agents do.","status":"reported","action":"Confirm each agent in production logs its instructions and actions somewhere a reviewer can read them.","companyLabel":"Public edition"},{"slug":"senate-ai-systems-transparency-act","kind":"rule","source":{"title":"Senators Coons, Britt, Schatz, and Lankford introduce bipartisan AI safety and transparency legislation","url":"https://www.coons.senate.gov/news/press-releases/senators-coons-britt-schatz-and-lankford-introduce-bipartisan-ai-safety-and-transparency-legislation","date":"2026-09-24"},"headline":"Four senators want AI labs to publish model cards, with the FTC enforcing","whatChanged":"On September 24, Senators Chris Coons, Katie Britt, Brian Schatz, and James Lankford introduced the AI Systems Transparency Act. Covered AI companies, defined by size criteria in the bill, would publish model cards and describe their safeguards for child safety, mental health, privacy, cybersecurity, disaster risk, and loss of control, updated with each new model. The FTC would enforce it for open and closed models alike. The same day, attorneys general from 23 states and D.C. asked Congress for federal AI safety rules.","whyItMatters":"The bill regulates the companies that build models, and it has only been introduced, with midterms in November. For buyers, it points at a practical gap. Model cards and safeguard descriptions are the documents a company needs when it decides whether to trust a model with its data, and today they are voluntary and uneven across providers. If the bill moves, vendors would have to publish them. Until then, buyers have to ask.","issuer":"U.S. Senate","glance":"Four senators introduced the AI Systems Transparency Act on September 24; the FTC would enforce new model disclosure rules.","status":"proposed","action":"Ask each AI vendor you use for its current model card and data terms, and file them with the approval.","companyLabel":"Public edition"},{"slug":"reported-frontier-ai-standards-body","kind":"rule","source":{"title":"Google, OpenAI, Anthropic Plan Frontier AI Standards Body","url":"https://www.bankinfosecurity.com/google-openai-anthropic-plan-frontier-ai-standards-body-a-32926","date":"2026-09-24"},"headline":"Google, OpenAI, and Anthropic reportedly plan their own industry testing body","whatChanged":"The Information reported on September 24 that Google, OpenAI, and Anthropic are planning a self-regulatory group called the Standards Authority for Frontier AI. It would set common standards for assessing frontier models and run benchmark tests. The plan borrows from FINRA, which is registered with the SEC. It could launch at the end of 2026 or early 2027 and would likely be funded mostly by industry. Unless it registers with a government agency, it would have no legal authority.","whyItMatters":"If it forms, its test results could become something buyers cite when they approve a model, much like SOC 2 reports for other software vendors today. For now it is only a report. It changes nothing about a company's own job of vetting AI tools before customer-facing use, and a standard funded by the labs it grades deserves some skepticism when it arrives. Teams that approve AI tools would still run their own checks.","issuer":"The Information, reported","glance":"Google, OpenAI, and Anthropic are reportedly planning a FINRA-style industry body to set testing standards for frontier models.","status":"reported","companyLabel":"Public edition"}],"older":{"id":"2026-09-21","publishedLabel":"Week of September 21, 2026","title":"AI agents get cheaper as new disclosure rules land","isLatest":false},"newer":{"id":"2026-10-05","publishedLabel":"Week of October 5, 2026","title":"Cheaper models, gated releases, and new California AI laws","isLatest":true}},"session":null}