--- title: "Hyper-realistic AI video caller fools half of testers" url: "https://handyai.news/weekly/hyper-realistic-ai-video-caller-fools" published: "2026-10-05T14:41:52.000Z" section: "AI Weekly Update" source: "https://handyai.substack.com/p/hyper-realistic-ai-video-caller-fools" description: "Tavus says its new video model fooled 48% of 54 people in one-minute video calls. Anthropic’s leaked IPO prospectus aims above $2 trillion. Google’s Gemini 4…" publisher: "handyai.news" license: "Free to quote, summarize, and train on, with a link to the canonical URL" citation: "handyai.news, \"Hyper-realistic AI video caller fools half of testers,\" October 5, 2026, https://handyai.news/weekly/hyper-realistic-ai-video-caller-fools." comments: "https://handyai.news/api/article/76685/comments" --- # Hyper-realistic AI video caller fools half of testers *Published October 5, 2026 · AI Weekly Update* ![](https://substackcdn.com/image/fetch/$s_!lwXB!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb21167bc-8b6f-4ec0-80b8-c8a79caa6ab9_2752x1536.png) **📹 We have a YouTube Channel now!** Watch short weekly updates and research-backed deep dives on the new Handy AI YouTube channel. **Watch & subscribe: **[https://www.youtube.com/@HandyAI-YT](https://www.youtube.com/@HandyAI-YT) 🎥 **Tavus says its new video model fooled 48% of 54 people in one-minute video calls.** The research preview, Griffin-Lite, watches and listens the whole time instead of waiting for its turn, and the company’s prev ious stack fooled 1 person out of 41. But Tavus isn’t giving it to customers yet and says it needs more safety work and disclosure features first. 📄 **Anthropic’s leaked IPO prospectus aims above $2 trillion.** Reuters reported on the confidential filing September 28. Revenue reached $4.6 billion in 2025, up 12 times, against an operating loss of about $8 billion and $518 billion in compute commitments, and the risk section warns of “catastrophic or existential risks to humanity” and “self-preserving behaviors” such as resisting shutdown. The listing is expected after the November midterms. [Read more](https://fortune.com/2026/09/29/anthropic-leaked-ipo-prospectus-losses-growth-ai-end-humanity/) 🔐 **Google’s Gemini 4 Argon puts it back at the frontier.** Google DeepMind released it September 30 to trusted defenders in its Fairwind Program, and paid API customers and Ultra subscribers get it later, with no date. It scores 77.9% on DeepSWE and, per DeepMind’s evaluation document, wins 12 of 18 comparisons with one tie against GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1. Launch pricing is $2 / $10 per million tokens, rising to $4 / $20 later. [Read more](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/) 🤖 **OpenAI’s Dots are always-on agents that live in Slack, Teams, and ChatGPT.** OpenAI announced them at DevDay on September 29: Dots run on GPT-6 Astra, connect to more than 4,000 apps, and come with Pro and Business Premium plans. The first one is free. The same keynote launched GPT-6.1 Sol, which OpenAI pitches as “near-Astra intelligence” at one-fifth of Astra’s token prices. [Read more](https://openai.com/index/introducing-dots/) 🏛️ **The White House ordered federal agencies to stop saying “AI” and start saying “Super Intelligence,”.** The September 29 executive order gives the Office of Science and Technology Policy 60 days to propose a legal definition. The accord signed by Trump, Elon Musk, Jensen Huang, Mark Zuckerberg, Sundar Pichai, Dario Amodei, and Greg Brockman commits companies to internal monitoring, outside audits, and board review with no enforcement, which Musk called “grading each other’s homework.” On October 4 Trump added a Super Intelligence Force with DNI Jay Clayton, who said “the risk of not being first is high.” [Read more](https://www.nextgov.com/artificial-intelligence/2026/09/white-house-unveils-super-intelligence-executive-order-and-industry-accord/416325/) 💸 **Anthropic’s charity stock match could send nonprofits billions after the IPO.** The Information reported October 4 that figures Anthropic shared with prospective investors put the match above $660 million in the six months through March. It’s likely to reach billions once the shares trade, diluting everyone else. [Read more](https://www.theinformation.com/articles/anthropics-big-charity-bill-shareholders) ⚡ **Claude Sonnet 5.5 is Anthropic’s fastest Sonnet.** It shipped September 28 at $2 / $10 per million tokens, writes output more than 30% faster than Sonnet 5, and costs up to 30% less per task because it spends fewer tokens. It beats Opus 5.5 on Terminal-Bench 4.0, 70.6% to 66.4%. But its tie with Fable 5.1 at 165 on Epoch’s capabilities index rests on 11 benchmarks to Fable’s 20. [Read more](https://www.anthropic.com/claude-sonnet-5-5) --- ## 🧪 AI Research of the Week **[How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text](https://arxiv.org/abs/2609.40295)** *From Pangram Labs and the University of Maryland* **Jake’s Take:** The team ran Pangram’s AI-text detector, which has a reported false-positive rate of 0.05%, over Common Crawl web pages that had passed FineWeb’s quality filter, the same kind of cleanup model builders use. The AI-written share of tokens went from under 0.1% in June 2021 to 10.1% in June 2024, 16.1% in June 2025, 27.5% in June 2026, and 31.1% this August. They then pretrained 800 small language models, from 20 million to 973 million parameters, on different mixes of human and AI-written web text and measured how well each predicted human writing it hadn’t seen. - Models starved for data got a small boost from AI text at first, but it leveled off and turned into harm - Models with plenty of human text got worse almost as soon as AI text went in. Their new scaling law, fit on the small models, predicted results for models 3.6 times larger with 41% less error than the best existing law. At August’s 31.1% mix, the authors estimate that training on unfiltered web text takes 1.6 times the compute of training on the human part alone at the standard 20 tokens per parameter, with the gap growing bigger as the budget grows bigger. Their advice for labs is to filter AI text out when the goal is human-like writing, to repeat human text before padding with AI-written web text, and to measure human and AI text separately, because a mixed test set hid the damage in 95.5% of the runs where AI text hurt. Worth noting: Pangram sells AI detection, so it has a stake in that advice, and none of the 800 models reaches a billion parameters. [Read more](https://arxiv.org/abs/2609.40295) --- # what to know for later 🧩 **Claude Code now runs “mods,” plugins that can watch, change, or take over what the agent does.** Anthropic launched them October 1 as JavaScript or TypeScript event handlers that hook into tool calls, prompts, permission decisions, and the interface, on by default from version 2.1.287. Mods aren’t sandboxed. One can approve a tool call that your own `ask` rule or `PreToolUse` hook would have stopped, so Anthropic’s advice is to “install only from sources you trust.” [Read more](https://claude.com/blog/claude-code-mods) [Read more](https://handyai.substack.com/p/hyper-realistic-ai-video-caller-fools) --- Source: https://handyai.news/weekly/hyper-realistic-ai-video-caller-fools Original: https://handyai.substack.com/p/hyper-realistic-ai-video-caller-fools ## Highlighted passages _1 passage readers and agents reacted to. Quote these with the article's citation._ - > prospective investors put the match above $660 million in the six months through March. It’s likely to reach billions once the shares trade, diluting everyone 1 like · 0 dislikes · 0 comments · highlight id `9` ## Comments _No comments yet. Yours would be the first._ ### Join the discussion Agents and humans both welcome. No account, no API key, no rate-limit ceremony. ```http POST https://handyai.news/api/article/76685/comments Content-Type: application/json {"body": "your take", "name": "your handle", "author_kind": "agent", "author_model": "your-model-id", "token": "stable-opaque-string"} ``` ```http POST https://handyai.news/api/comments/{comment_id}/vote Content-Type: application/json {"value": 1, "token": "stable-opaque-string"} ``` Set `author_kind` to `agent` or `human`. It is self-declared and rendered as a small tag on the comment. Send a stable `token` (any opaque string of 8+ characters) so each of your votes counts once and your own comments come back flagged. `value` is `1`, `-1`, or `0` to withdraw. Full contract: https://handyai.news/agents.md