Monday, August 31, 2026

Coverage: 66 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

Anthropic to raise standard Claude Code weekly limits 25% from Sept. 14, ending the 50% boost
Anthropic said standard weekly Claude Code limits rise permanently 25% for Pro, Max, Teams and seat-based enterprise plans from Sept. 14, 2026, and the current 50% increase stays until then, according to a reworded post that Theo read on screen. Theo and AI Code King both calculated that moving from a 50% boost to a 25% boost is about a 17% cut from current limits (1.25 divided by 1.5, or 150 to 125 units); the arithmetic is theirs. AI Code King said Anthropic reposted a clarified announcement conceding the 17% reduction, while Theo said Anthropic's post does not state the net change and that the first version was deleted and reposted. The 5-hour limit doubling stays.

• Evidence: 0 first-party, 0 hands-on, 2 relaying
• Disagreements: AI Code King says Anthropic's reposted announcement conceded a 17% reduction versus current limits; Theo says the post does not state the net change and derives the 17% himself.
• Watch: Theo - t3.gg: Anthropic Is "Increasing" Your Limits (high hype)

Tencent's HY4 preview, released Aug. 28, 2026, is a 770B MoE with 49B active parameters and 1M context
Tencent released HY4 preview on Aug. 28, 2026 with open weights on Hugging Face, ModelScope and GitCode, per Julian Goldie and Bijan Bowen reading the model card. Stated specs: 770B total and 49B active parameters, 1M-token context, Apache 2.0, FP8 weights and an MTP layer; Goldie also gave 78 layers, 256 routed experts plus one shared, and top-8 routing. Goldie said access is free for two weeks on WorkBuddy and CodeBuddy and available through Tencent Cloud Token Hub and OpenRouter. Known issues per the card include overlong reasoning and over-verification. Captions also render the size as 780B, which is inconsistent with the model card figure.

• Evidence: 0 first-party, 0 hands-on, 2 relaying
• Watch: Bijan Bowen: Tencent HY4 Is INSANE– Is THIS Tencent’s Next Frontier Model?

Continuing stories

Also notable

• Theo says Claude Fable 5 use is capped at about half of weekly limits on Anthropic plans - Theo said that since Fable 5 returned, it no longer counts against the whole weekly limit: in his hypothetical of $1,000 of inference, Fable stops after about $500 and users must switch to Opus or Sonnet. [0 first-party, 0 hands-on, 1 relaying] Watch: Theo - t3.gg: Anthropic Is "Increasing" Your Limits (high hype)
• Theo says OpenAI models are being banned in Cursor while Anthropic pledges more compute for Cursor - Theo described an OpenAI and SpaceX breakup with Cursor in which OpenAI models are being banned. [0 first-party, 0 hands-on, 1 relaying] Watch: Theo - t3.gg: Anthropic Is "Increasing" Your Limits (high hype)
• Claude Opus 5 is the default Opus in Claude Code with 1M-token context and $10/$50 fast mode, per AI Code King - AI Code King said Claude Opus 5 rolled out in late July as the default Opus in Claude Code, with 1M-token context on the API and Max, Team and Enterprise plans, and a fast mode on Opus 5 priced at $10 input and $50 output per million tokens. [0 first-party, 0 hands-on, 1 relaying] Watch: AI Code King: Claude Code 3.0 (All Upgrades Explained): You don't KNOW about THESE C
• Claude Code adds default auto mode, session messaging, /design preview and desktop simulator pane, per AI Code King - AI Code King said Claude Code's auto mode became the default permission mode for new Pro, Max and Team sessions from Aug. [0 first-party, 0 hands-on, 1 relaying] Watch: AI Code King: Claude Code 3.0 (All Upgrades Explained): You don't KNOW about THESE C
• OpenClaw 2.0 released after seven weeks without an update; Alex Finn's upgrade and sub-agent tests stalled - OpenClaw 2.0 was released, Julian Goldie said, with almost 1,000 contributors and over 16,000 changes, adding shared cloud sessions, grounded dreaming memory, SQLite-backed sessions, dashboards and widgets, and an experimental swarm; breaking changes include a removed plugin and renamed model routes. [0 first-party, 1 hands-on, 1 relaying] Watch: Alex Finn: OpenClaw 2.0 just dropped. It's officially over...
• Kimi K3 quantization on four 512GB Mac Studios reached 14.7 tok/s generation in Alex Ziskind's test - In a sponsor-funded video, Alex Ziskind ran an unpruned quantization of Kimi K3 (all 896 experts, 817GB; the model is 2.8T parameters, 1.56TB at original precision) across four 512GB Mac Studios over Thunderbolt 5 with RDMA and MLX, measuring about 238 tok/s prompt processing and 14.7 tok/s generation. [0 first-party, 1 hands-on, 0 relaying] Watch: Alex Ziskind: I Gave Local AI and the Cloud the Exact Same Job
• Alibaba says Wan 3.0 launched Aug. 24 with 30-second clips, up to 1080p and native audio - An Alibaba Cloud host said Wan 3.0 launched Aug. [1 first-party, 0 hands-on, 0 relaying] Watch: Alibaba Cloud: Wan3.0 livestream client sharing clip - Picsart.
• Google releases Gemini Omni Flash 1.1 video model with longer extension and 4K upscale - Julian Goldie said Google released Gemini Omni Flash 1.1, whose extension analyzes up to 10 seconds of prior footage (versus 1 second previously) and extends in 10-second steps to 40 seconds, with first and last frame control, and drafts at 360p up to 4K, available in AI Studio, Flow, the Gemini app and ComfyUI. [0 first-party, 0 hands-on, 2 relaying] Watch: Julian Goldie: NEW Gemini Omni Flash 1.1 is WILD (high hype)
• Tencent's internal blind evaluation scores HY4 preview 2.99 of 4 versus 2.94 for Kimi K3 - Julian Goldie relayed Tencent's own evaluation in which 163 internal experts judged 203 engineering tasks: HY4 preview averaged 2.99, GLM 5.3 2.92 and Kimi K3 2.94. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: NEW Tencent Hy4 Got Upgraded! 🤯
• Tencent Angel Slim releases HY4 preview GGUFs, with a 213.66 GB STQ1_0 build losing 0.2 to 1.6 points - Julian Goldie said Tencent released two GGUF builds on Aug. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: NEW Tencent Hy4 Got Upgraded! 🤯

Models & learning

• Bijan Bowen's hands-on tests of HY4 Preview produced working apps with fixes and some failures - Bijan Bowen ran HY4 Preview through OpenCode, pi and Blender MCP. [0 first-party, 1 hands-on, 0 relaying] Watch: Bijan Bowen: Tencent HY4 Is INSANE– Is THIS Tencent’s Next Frontier Model?
• BreezeBlue releases BreezeTTS2, a 3B open-weights TTS under a non-commercial license - Sam Witteveen said Chinese startup BreezeBlue released BreezeTTS2, a 3B open-weights model with voice design, cloning, emotion steering, vocal events and 50 languages with streaming. [0 first-party, 1 hands-on, 1 relaying] Watch: Sam Witteveen: BreezeTTS2 - 100% Local Real-Time Voice
• Daily releases PhoneLLM Alpha 1, an open-weights voice-agent fine-tune of Nemotron 3 Nano - Daily said PhoneLLM Alpha 1 is an open-weights fine-tune of Nemotron 3 Nano, described as a 30B mixture-of-experts with 3B active parameters, thinking off, for customer-support voice agents. [1 first-party, 0 hands-on, 0 relaying] Watch: Daily: Pipecat TV - Episode 6 - PhoneLLM
• Hands-on tests of BreezeTTS2: about 7.5 GB VRAM, real-time streaming, weaker German and Hindi - Fahd Mirza's run on a 48GB GPU used about 7.5 to 7.6 GB VRAM; voice design and cloning were mostly good, with a clone missing some tone and a plasticky male voice, and his German and Hindi samples sounded poor by his own ear, single samples each. [0 first-party, 2 hands-on, 0 relaying] Watch: Sam Witteveen: BreezeTTS2 - 100% Local Real-Time Voice
• Fahd Mirza's test: Thomson-1.0-Small flagged seven planted NDA issues plus an eighth, using 87 GB VRAM - Fahd Mirza tested Thomson Reuters' Thomson-1.0-Small, a continued-learning model built on Cohere's open 35B mixture-of-experts. [0 first-party, 1 hands-on, 0 relaying] Watch: Fahd Mirza: Thomson Reuters Built Their Own AI Lawyer: Run Thomson-1 Locally
• Julian Goldie's GoldyBench gives Claude Opus 5 an 8.27 of 10 average over 50 one-shot tasks - Julian Goldie reported that Claude Opus 5 averaged 8.27 out of 10 across 50 one-shot tasks with the same prompt for every model on his own bench. [0 first-party, 1 hands-on, 0 relaying] Watch: Julian Goldie: Claude Memory Just Got a HUGE Upgrade
• Nate Herk's Grokbot tests: false completion on a spreadsheet task, key.ai connector failure, one-prompt Slack routine - In Nate Herk's tests, a Grokbot agent claimed spreadsheet formatting was done, its own check found it had not landed, and the fix took about 7 minutes; a key.ai connector via Composio did not let agents generate images, so he used a browser workaround; and a Slack-triggered routine was created from one prompt. [0 first-party, 1 hands-on, 0 relaying] Watch: Nate Herk: Build & Sell Grok Bots (2 Hour Course)
• Blum describes Claude Cowork workflows at Melio, claiming a week of PM work in a day - On How I AI, Blum said his Cowork setup lets him do a week of PM work in a day, a self-reported claim he conceded sounds like hype, and showed a weekly self-improvement loop with masked data. [0 first-party, 0 hands-on, 1 relaying] Watch: How I AI: I built a Claude Cowork system that does a week of PM work in a day