Busy week: GPT-6.2 Astra, Gemini 4 Carbon, Fable 5.2 and two batches of open weights.
Just out
GPT-6.2 Astra and Sol: OpenAI released both today. They’re available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API. Astra costs $8/$40 per million input/output tokens, 20% less than GPT-6 Astra. It scores 5.2 points higher than GPT-6 Astra on Agents’ Last Exam at under a third of the cost per task. Sol keeps GPT-6.1 Sol’s API prices. GPT-6.2 Astra and Sol are now available - Announcements - OpenAI Developer Community +2
LLMs
Inkling 2 Small: Thinking Machines released the weights of a 594B MoE with 17B active parameters. The NVFP4 version needs two B300s or four H200s.
Gemini Decision 1: Google now has its own Jev.
Gemini 4 Carbon: out in the paid API and in Gemini for Ultra subscribers. $2/$10 per million input/output tokens until November 30, then $4/$20.
Fable 5.2: generates responses up to 30% faster than 5.1. API price is $8/$40 per million input/output tokens.
Kimi K3.1 Pro weights: Moonshot released them and swapped Gated MLA for Gated DSA to handle long context more efficiently.
Generative models
MiniMax H3-Image: open weights, native 4096×4096 in six steps, up to 16 references.
FLUX 3 Max promo: fal offered free text-to-video and image-to-video with sound, up to 15 seconds at 720p.
Gemini Omni 1.2 Flash: video generation and editing with sound, up to 15 seconds. API price is about $0.07 per second at 720p, input tokens billed separately.
Other
Codex retro console: OpenAI released the ModRetro Chromatic for $200. You can build your own games in Codex and flash them to a cartridge.
@Alibaba_Qwen@youdotcom 1M context with native video input is the interesting part. Can I drop a full 10-minute screen recording in and get a usable bug report back?
@LyalinDotCom Nice, but the CLI can’t use subscription quota for images and video, only the API. So the $10 is gone after about 20 minutes of generation.
@HaralAfif Zero code edits on the build itself. Same prompt, one run, what you see is what it made. The only manual part was the video edit, cutting it down and adding captions.
Same task for every new AI model: a 30-second video about Cyprus.
Fable 5.2 built the 3D island and edited the film. Top: clips by Grok Imagine. Bottom: Kling 4.0 with Gemini start frames.
Grok invents Limassol. Kling + Gemini looks like real drone footage.
@euronews The real story isn’t the petition. A $4T+ company had no succession plan for a CEO who’s been in the chair for 33 years. That’s on the board, whatever happens with Jensen.
@nathie The risk isn’t the hardware, it’s “a version of visionOS.” Vision Pro had a whole OS and barely any apps worth opening. Glasses need three things that work perfectly, not a platform.
@IamEmily2050 Local Gemma + big model in the cloud makes sense for summaries and small edits. For long agent runs it won’t save you, the heavy steps still burn the expensive quota.
Starship has been on my mind, so I used it to test MiniMax’s new H3-Image.
Two reference photos and one prompt gave me a 360° launch pad panorama in native 4K, in 20 seconds. The model takes up to 16 reference images, and the weights are open. The launch clip is H3 video, made from a single frame of the panorama.
Chinese labs still trail the US frontier companies by about 6 to 12 months on LLMs, but on image and video generation they seem to be pulling ahead.
@AiXsatoshi Funny part: a second Mac over Thunderbolt RDMA only got about 1.7x, and an NVIDIA box over plain 10GbE got about 7x. Kind of changes the math on Max vs Ultra if you can pair a Max with a GPU box for prefill.
@SamJWasserman Clean setup. What do you actually run across all those Sparks? One big model split over several boxes, or a separate model per box for different jobs?
@AmateurMika@omarchy@JoshuaSWarren The real test is day 30, not day 2. How’s battery life, sleep/wake and the internal speakers? Those are usually what break Linux on new Apple Silicon.
@PawelHuryn Half the Mac mini crowd isn’t buying RAM. They want macOS: iMessage, Apple Notes, Shortcuts and other stuff agents can’t reach from Linux. For pure compute, agreed, a used mini PC wins.
@ashxhart 3.2–3.5x across 24 streams is a batching win. What does a single stream look like on the M5 Ultra? That’s the number for anyone running one agent locally.