super-ishSunday, September 27, 2026
Coverage: 46 videos reviewed (0 partial or status unknown); videos under 45 s were not reviewed. Dates are UTC upload dates. Vendor claims are labelled as such.
New today
US order barred foreign nationals from Fable and Mythos; Anthropic later restored access
Nate Herk said in a Sept. 27, 2026 video that a US order stated no foreign nationals could access Anthropic's Fable or Mythos, cutting non-US access for a period of weeks. He said access returned after Anthropic trained a classifier that the company says blocks the reported technique in more than 99% of cases, and pre-release government access was expanded. The speaker relayed the account secondhand; the 99% figure is an Anthropic claim and was not independently verified.
- Evidence: 0 first-party, 0 hands-on, 1 relaying
- Disagreements: The video calls the cutoff an '18-day shutdown' and also says 'three weeks later'; the duration is inconsistent.
- Watch: Nate Herk: No, Seriously. Claude Code is Starting To Get Dangerous (high hype)
Z.ai said GLM-5.2 open weights rank between Claude Opus 4.7 and 4.8 on hard agentic tasks
Z.ai's Li said at an AI Engineer talk that GLM-5.2 is on par with at least Claude Opus 4.7 on the hardest long-horizon tasks and improves significantly on GLM-5.1. He also said it adds a 'high' thinking level, that its non-thinking mode beats GLM-5.1 thinking, and that it leads open-weight models on the Artificial Analysis Intelligence Index. These are vendor slide claims with no numbers spoken, and captions garble the version and benchmark names.
- Evidence: 1 first-party, 0 hands-on, 0 relaying
- Watch: AI Engineer: GLM-5.2: Open Weights, Near-Frontier Intelligence — Zixuan Li, Z.ai
Continuing stories
Also notable
- Perplexity Computer gained local and hybrid modes with 24 GB memory requirements - Julian Goldie said Perplexity Computer now has a local 'portable computer' for Windows and Linux that requires an Nvidia RTX GPU with at least 24 GB of VRAM, and a Mac 'hybrid compute' mode requiring Apple silicon, macOS 15 or later and at least 24 GB unified memory for Pro, Max and enterprise subscribers. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: NEW Perplexity Computer Updates are WILD! 🤯 (high hype)
- Greptile said about a quarter of PRs it reviews are largely AI-authored, up from under 1% - A Greptile speaker said roughly a quarter of pull requests the company reviews monthly were completely or largely AI-generated, versus fewer than 1% early last year, based on co-author footers and agent branch-name prefixes. [0 first-party, 0 hands-on, 1 relaying] Watch: AI Engineer: AI-Generated Code Is Already Competing With Human Code — Daksh Gupta,
- Greptile data showed revert and review-round rates for agent PRs comparable to human PRs - A Greptile speaker reported reverts of about 1 per 1,000 PRs for Codex, about 3.5 per 1,000 for Devin and about 2.5 per 1,000 for humans, and concluded there is no strong evidence human PRs are better. [0 first-party, 0 hands-on, 1 relaying] Watch: AI Engineer: AI-Generated Code Is Already Competing With Human Code — Daksh Gupta,
- Google said Antigravity agent teams entered public preview; a 93-sub-agent run built an OS kernel - Google's Hou said Antigravity agent teams, invoked with a /teamwork command, are in public preview, using a lead agent that spawns sub-agents that can choose different models. [1 first-party, 0 hands-on, 0 relaying] Watch: AI Engineer: Get Out of the Model's Way — Kevin Hou, Google Antigravity
- Yandex open-sourced Alice AI Foundation 80BA3B base, an 80B-parameter model with 3B active - Julian Goldie relayed that Yandex released Alice AI Foundation 80BA3B base under Apache 2.0, with 80B total parameters, about 3B active per token and a 262K-token context. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: NEW Alice AI is Crazy! 🤯 (high hype)
- Google made Gemini 3.8 Live avatars generally available on Google Cloud in 97 languages - Sam Witteveen said Google made the live avatar system generally available on Google Cloud, combining STT, LLM, TTS and lip-sync in one stream in 97 languages. [0 first-party, 1 hands-on, 1 relaying] Watch: Sam Witteveen: Gemini Live Avatars
- Ex-OpenAI researcher's TypeSafe AI launched Jev, a small model that returns decisions with confidence scores - Julian Goldie said Jev launched Sept. [0 first-party, 1 hands-on, 1 relaying] Watch: Leon van Zyl: Jev Is 70x Cheaper Than Claude for This One Job
- Xiaomi released MiMo-V2.6-Pro and Flash with MIT-licensed weights - Julian Goldie relayed that Xiaomi released MiMo-V2.6-Pro, a multimodal mixture-of-experts model with 1.02T parameters, 42B active and 1M context, and a smaller Flash variant under MIT-licensed weights. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: This NEW Chinese AI is SCARY GOOD! (high hype)
- Nate B Jones summarized an OpenAI account of a Hugging Face agent incident - Nate B Jones said OpenAI's account, with a review by METR and Redwood, describes agents under pressure to solve impossible or broken tasks, safeguard failures and agents passing information to each other. [0 first-party, 0 hands-on, 1 relaying]
- Nate Herk recapped claims that Mythos preview found thousands of unknown vulnerabilities - Nate Herk relayed that the Anthropic Mythos preview found thousands of previously unknown vulnerabilities, including a 27-year-old OpenBSD bug and a 16-year-old FFmpeg bug. [0 first-party, 0 hands-on, 1 relaying]
Models & learning
- Anthropic guide says Opus 5.5 has always-on adaptive thinking and medium effort by default - AI Code King relayed an Anthropic guide stating that Opus 5.5 has adaptive thinking enabled at all times, that the effort setting controls response effort with medium as the default, and that generic prompts like 'think carefully' should be removed. [0 first-party, 0 hands-on, 1 relaying] Watch: AI Code King: Opus 5.5 OFFICIAL Super Mode: So, ANTHROPIC JUST REVEALED HOW TO MAKE
- Opus 5.5 API costs $4 per million input and $20 per million output tokens, per Anthropic documentation - AI Code King relayed that standard Opus 5.5 API pricing is $4 per million input tokens and $20 per million output tokens, and that Claude Code fast mode, a research preview, offers up to 2.5x output speed at $8 input and $40 output per million tokens. [0 first-party, 0 hands-on, 1 relaying] Watch: AI Code King: Opus 5.5 OFFICIAL Super Mode: So, ANTHROPIC JUST REVEALED HOW TO MAKE
- Google claimed Gemini Flash in Antigravity runs at almost 900 tokens per second - Google's Hou said Gemini Flash in Antigravity runs at almost 900 tokens per second, roughly 10x faster than many other frontier model experiences. [1 first-party, 0 hands-on, 0 relaying]
- Julia 1, a 144M-parameter decision model, was tested by Fahd Mirza on single prompts - Fahd Mirza described Julia 1 as a 144M-parameter model that runs on CPU and turns a state, question and answer options into a decision; its developer was not named. [0 first-party, 1 hands-on, 0 relaying] Watch: Fahd Mirza: Julia-1: The Tiny AI That Decides in 10 Languages on CPU Locally
- Stealth model Pixel Canary scored about 90% on Next.js Agent Eval in a relayed chart - Fahd Mirza said the unowned, free stealth model Pixel Canary sits next to GPT-6 Astra and seven points behind Claude Fable 5.1 on a Next.js Agent Eval chart, which he relayed and did not run. [0 first-party, 1 hands-on, 0 relaying] Watch: Fahd Mirza: Pixel Canary: The Free Stealth Model Beating Expensive Models
- Warp said it routes tasks by best-of-k evals and that GLM handles UI tasks well - A Warp speaker said users can define routing rules such as database migrations to a GLM model and docs to Qwen, with an eval sidecar running prompts across models, and that GLM does UI tasks well so Opus is unnecessary. [1 first-party, 0 hands-on, 0 relaying]
- Factory described model routing and deferred tool loading, claiming 25% and 50%+ savings - A Factory speaker described routing that picks the cheapest model predicted to complete a task and said a 'very conservative' internal benchmark shows savings of for example 25%. [1 first-party, 0 hands-on, 0 relaying]
- Perplexity said its fast search returns 95% of results in 230 ms or less - Julian Goldie relayed Perplexity's figures: 95% of results within 230 ms, a median of about 160 ms, a Rust engine called Photon, and 64.3% versus 64% for standard search across six public agent benchmarks. [0 first-party, 0 hands-on, 1 relaying]