Saturday, September 12, 2026

Coverage: 41 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

Creators report mixed results from GPT-6 Astra in Codex and ChatGPT Work
Several creators described their own use of OpenAI's GPT-6 Astra in videos posted Sept. 12, mostly favorably and without controlled comparisons. Cole Medin said Astra beat Fable 5.1 in most of a week of his own testing and needed less intent-explaining than Opus 5. Alex Finn, in a sponsored video, called Astra the fastest and best computer-use model and said one task saved about 4 hours. Julian Goldie said Codex with Astra built a motion-design tool in about 7 minutes. Medin said benchmarks looked roughly equivalent between Astra and Fable 5.1. In a Goldie livestream, one speaker said Astra had gotten worse in Codex; the remark was anecdotal, with no comparison.

• Evidence: 0 first-party, 0 hands-on, 3 relaying
• Disagreements: Medin, Finn and Goldie report favorable results; one speaker in a Goldie livestream said Astra had gotten worse in Codex. All accounts are subjective and none share task-level data.
• Watch: Cole Medin: GPT-6 Astra Just Made AI Software Factories Real (Here's How to Run On; Alex Finn: ChatGPT Work with GPT 6 Astra just blew my mind (high hype)

Goldie and Dylan Davis relay Sept. 3 GPT-6 Astra release and its per-token price
Julian Goldie said OpenAI released GPT-6 Astra on Sept. 3 with computer-use ability. Dylan Davis said Astra and Claude Fable 5.1 were released about a week before his video and share $10-in, $50-out per-token pricing. Davis also said Astra is usually 8 to 9 times cheaper per task than Fable 5.1, without giving task details.

• Evidence: 0 first-party, 0 hands-on, 2 relaying
• Watch: Dylan Davis: I Stopped Choosing Between ChatGPT and Claude. Here's the Setup; Julian Goldie: GPT 6 Astra + Hermes Agent is Crazy Good! 🤯 (high hype)

Continuing stories

Also notable

• Finn walks through ChatGPT Work plugins, projects, routines and cloud computer option - Alex Finn, in a sponsored video posted Sept. [0 first-party, 1 hands-on, 0 relaying] Watch: Alex Finn: ChatGPT Work with GPT 6 Astra just blew my mind (high hype)
• Speakers differ on GPT-6 Astra cost: cheaper per task, costly by default, capacity-heavy - Speakers in Sept. [0 first-party, 1 hands-on, 2 relaying] Watch: Bart Slodyczka: I Tested OpenAI's New Cloud Agents... What You Need To Know
• OpenAI released an Agents API exposing the Codex harness as cloud agents - Bart Slodyczka said in a sponsored video posted Sept. [0 first-party, 1 hands-on, 0 relaying] Watch: Bart Slodyczka: I Tested OpenAI's New Cloud Agents... What You Need To Know
• DeepSeek V4.1 Flash offered free for about two weeks via WorkBuddy and Token Harbor - Two channels reported temporary free access to DeepSeek V4.1 Flash. [0 first-party, 1 hands-on, 1 relaying] Watch: AI Code King: FULLY FREE Deepseek V4.1 Flash Coder: This SHOULDN'T BE FREE! (+My Des
• Perplexity open-sourced Lily, a Rust and Metal engine for one local model on Macs - Julian Goldie said Perplexity open-sourced Lily, a Rust and Metal inference engine built for one model (about 35B parameters, about 3B active) and one chip family, with a working builder demo as the open-sourced part. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: Perplexity Just Open-Sourced Its Local AI Engine
• Manolo Remiddi runs Qwen 3.8 quant at about 45.6 tokens/s on 16GB GPU - Manolo Remiddi measured about 45.6 tokens per second for a 400-token story on a 16GB RTX 5060 Ti with a Qwen 3.8 MTP quant at 64,000-token context, using 15.1 of 15.9 GB; the speed was read from the agent's own report after one prompt. [0 first-party, 1 hands-on, 0 relaying] Watch: Manolo Remiddi: 16GB Is All You Need for Serious AI
• DeepMind released AlphaGenome Atlas of precomputed effects for about 9 billion DNA changes - Julian Goldie said DeepMind released AlphaGenome Atlas, a dataset of roughly one petabyte holding an AlphaGenome variant impact score for all single-letter DNA changes, about 9 billion, searchable through a website. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: Google Antigravity Just Changed Genomic Research (high hype)

Models & learning

• Two hosts report fast, usable front-end output from DeepSeek V4.1 Flash in agent harnesses - AI Code King built a fictional design-studio site and, with the Impeccable skill, an analytics dashboard with V4.1 Flash in WorkBuddy; the site worked and the dashboard's first render had a layout problem fixed after one feedback pass. [0 first-party, 2 hands-on, 0 relaying] Watch: AI Code King: FULLY FREE Deepseek V4.1 Flash Coder: This SHOULDN'T BE FREE! (+My Des
• Goldie relays Kimi K3 ranking first on Frontend Code Arena and long-horizon design - Julian Goldie said Kimi K3 ranked first on Frontend Code Arena, winning six of seven categories, and beat Fable 5 and GPT-5.6 in blind developer voting, with Fable 5 ahead in gaming. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: This Open-Source AI Agent Can Run for Days (high hype)
• Nex N2.5 family lists 35B, 397B and 1.6T models with vendor-reported scores - Julian Goldie said Nex N2.5 comes in three sizes: Mini at 35B parameters with vision and a free rate-limited route on OpenRouter, Pro at 397B and multimodal, and Max at 1.6T and text-only. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: New NEX N2.5 is WILD! ( FREE! ) 🤯 (high hype)
• OmniRoute open-source router lists 352 providers and claims about 1.47B free tokens a month - Julian Goldie said OmniRoute has over 61,000 GitHub stars and lists 352 providers, 150 or more with free options. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: Free Codex Is Absolutely WILD! (high hype)
• Qwen released Qwen Drive 1.0 4B; local test planned acceleration through a green light - Fahd Mirza said Qwen released Qwen Drive 1.0 4B, a vision-language model with a bird's-eye-view perception head and a planner that outputs a 5-second trajectory, in imitation-trained and RL-tuned versions. [0 first-party, 1 hands-on, 0 relaying] Watch: Fahd Mirza: Qwen-Drive-1.0-4B: Why You Still Can't Trust AI with Self-Driving Cars
• YuE2 open music model claims to beat Suno v6; test used under 8 GB - Fahd Mirza said the YuE2 model card and chart claim the 3B-parameter model outscores Suno v6 on a song-quality benchmark, generating an editable ABC-notation score and then 48 kHz audio. [0 first-party, 1 hands-on, 0 relaying] Watch: Fahd Mirza: YuE2 - Open Music Generation Model for Any Language Locally
• Non-uniform GSQ+RCO GGUF of Qwen 3.8 27B is said to match unquantized model - Manolo Remiddi said a non-uniform GSQ+RCO GGUF of Qwen 3.8 27B, which assigns each tensor its own quantization type under a size budget, is said to match the unquantized model. [0 first-party, 0 hands-on, 1 relaying] Watch: Manolo Remiddi: 16GB Is All You Need for Serious AI
• NVIDIA showed TensorRT Model Connect, a checkpoint-to-inference workflow, in a developer session - NVIDIA staff demonstrated TensorRT Model Connect, which they described as a feature of TensorRT rather than a new product, taking a Qwen3 0.6B checkpoint to inference in two commands on an RTX 5090. [1 first-party, 0 hands-on, 0 relaying] Watch: NVIDIA Developer: From Video to Voice: Build Faster with TensorRT Model Connect