Matthew Berman's 18-minute AI news roundup covers two frontier launches, smaller agent models, open-weight releases and image and video tools. It also opens with a serious NVIDIA story. The useful way to read this week is to separate confirmed availability from planned releases, company benchmarks and the host's own experience. Those are different kinds of evidence.
Fable 5.2 and Gemini 4 Carbon are available, but serve different cost and workflow needs. LongCat 2.5, Ripple 65B and MiniMax H3 Image have downloadable weights, while Inkling 2 Small's weight release was still pending on its official page when checked on 12 November. The NVIDIA board is investigating a reported security concern; it has not announced that any secrets were compromised.
Watch Matthew's Roundup
What Actually Shipped This Week?
This is a release map, not a quality ranking. "Open weights" means a model file can be downloaded; it does not by itself mean an unrestricted license or easy consumer-hardware setup. Availability and pricing below reflect official pages checked on 12 November 2026.
| Release | What it is | Status and original source |
|---|---|---|
| Claude Fable 5.2 | Frontier model for ambitious technical work | Available from Anthropic |
| Gemini 4 Carbon | Google's new frontier model | API access; product rollout |
| LongCat 2.5 | Large multimodal open-weight model | Weights and model card |
| Ripple 65B Preview | Dense coding and agent model | Preview and weight access |
| Inkling 2 Small | Thinking Machines' new model | Tinker access; weights pending |
| Muse Ember | Small model for repeated agent tasks | Meta API and Muse Code |
| GPT Image 2.6 | Image editing and GIF generation | ChatGPT, Codex and API |
| Gemini Omni 1.2 Flash | Conversational video generation | Google's release and access details |
| Seedream 5.5 | Image generation and multi-turn editing | ByteDance release |
| SWE-2 Flash | Fast specialist coding model | Devin access; limited weight program |
| MiniMax H3 Image | Open-weight image model | Weights under a custom license |
NVIDIA: What Is Confirmed, and What Is Not
On 8 November, Jensen Huang publicly acknowledged an extramarital affair and apologized. The NVIDIA board opened an independent investigation following reporting about a possible foreign-intelligence effort. It said it had no reason to believe that national security or NVIDIA secrets were compromised, while promising to examine the available information. On 10 November, independent directors named outside investigative counsel.
The video's suggestion that sensitive information might have leaked is a possibility being discussed, not an established finding. Neither the allegation about the other person's identity nor a breach should be repeated as fact. The board's process, not a video headline, is the reference point for what is known.
Fable 5.2 vs Gemini 4 Carbon: Different Buying Questions
Anthropic launched Fable 5.2 on 9 November for difficult, open-ended technical work. It reports gains in coding, reasoning and vision, up to 30% faster output than Fable 5.1, and an API price of $8 input / $40 output per million tokens. Anthropic still recommends Opus 5.5 for most coding and knowledge tasks. Fable and Mythos share an underlying model, but Mythos has a restricted verification route for certain research uses.
Google launched Carbon on 10 November, with paid API access and rollout to AI Ultra in Antigravity, Gemini and AI Mode. Its introductory API price is $2 input / $10 output per million tokens through 30 November, then $4 / $20 from 1 December. That makes the price comparison date-sensitive.
Matthew strongly prefers Fable in his own coding workflow. He shows an Excel-style browser app, a Fall Guys-style game and a Rube Goldberg simulation made with Carbon, noting real issues such as clipping and imperfect physics. His claim that Fable feels substantially better is a hands-on judgment, not proof that Carbon is worse on every task. The Excel build's reported time and token total came from asking the model about its own work; they are not a verified billing receipt. Our separate three-build review offers another practical comparison.
Open Weights and Specialist Coding Models
LongCat 2.5 is a large multimodal mixture-of-experts release. Its authors describe roughly 1.6 trillion total parameters with 48 billion active per token. Downloadable does not mean laptop-friendly: check the actual files, quantization options and memory requirements before planning a local deployment. Ripple 65B Preview is a smaller dense model with a claimed million-token context window. The 80.52 Terminal-Bench 2.1 result on its launch page is a vendor-reported benchmark, useful for choosing a test, not a guarantee for your repository.
The largest correction concerns Inkling 2 Small. Matthew encourages viewers to download its weights, but Thinking Machines' release says it launched through Tinker with closed weights while external safety evaluations continue. Its planned Apache 2.0 weight release should not be described as already complete. It has 594 billion total and 17 billion active parameters, which also makes casual local-running advice premature.
Muse Ember targets repetitive agent work such as tool calls, extraction and coding subtasks. Meta's efficiency chart uses its own proposed pricing and selected benchmark settings, so test it on a representative subtask rather than assuming it can replace a larger planning model. SWE-2 Flash is a similar specialist option for coding. Cognition says it started from Inkling-Small, not Inkling 2 Small, and is pursuing controlled weight access for partners and researchers. It is available in Devin, but not a general public weight download.
Image and Video: Four Different Workflows
GPT Image 2.6 adds animated GIF generation and editing, with OpenAI reporting roughly 50% faster generation on average. Sunburst prioritizes quality; Flare prioritizes speed. Image API work is still metered: OpenAI's model documentation lists token prices by modality, so the video's uncertainty about whether image models use tokens should not be read as free or flat-rate generation.
Gemini Omni 1.2 Flash focuses on short video clips with conversational editing; Google reports faster output and a lower price than 1.1. Seedream 5.5 emphasizes realistic images and iterative edits. Matthew calls its examples exceptional, but visual preference and consistency need testing on your own brand assets.
MiniMax H3 Image is a downloadable image model under a MiniMax community license, not an unrestricted open-source license. Its main repository is about 68 GB and also uses a shared text encoder. The separate 17.9 GB H3-Image-PE file is an optional prompt expander, not an 18 GB quantized copy of the full image model. Check the model card and license before spending time on a local install.
What Should You Try First?
For a hard technical project, scope one measurable task and compare Fable 5.2 with your existing model at the same acceptance criteria. For high-volume, bounded agent steps, test Carbon, Muse Ember or SWE-2 Flash where access and price fit; keep the expensive model for planning and review if it demonstrably improves the result. For local ownership, begin with the actual LongCat, Ripple or MiniMax model cards and hardware budgets, not the word "open" in a thumbnail. For visual work, compare a short real brief across GPT Image, Seedream and Omni, including editability, rights, render time and total cost.
Video Chapters
| Time | Segment |
|---|---|
| 00:00 | Intro |
| 00:15 | NVIDIA CEO story |
| 01:28 | Fable 5.2 |
| 04:31 | Sponsor |
| 05:35 | Fable 5.2 continued |
| 07:46 | Gemini 4 Carbon |
| 09:59 | LongCat 2.5 |
| 10:38 | Ripple 65B |
| 11:44 | Inkling 2 Small |
| 12:48 | Muse Ember |
| 13:52 | GPT Image 2.6 |
| 14:46 | Gemini Omni 1.2 |
| 15:13 | Seedream 5.5 |
| 16:06 | SWE-2 Flash |
| 16:48 | MiniMax H3 Image |
Sources, Corrections and Scope
This article uses the video and supplied transcript, plus the original company statements and model cards linked beside each claim. We checked their published status on 12 November 2026. The most important transcript corrections are Inkling 2 Small's pending weights, SWE-2 Flash's Inkling-Small base, and the MiniMax prompt expander's distinction from the full image model. Benchmarks and speed claims are attributed to their publishers; Matthew's demonstrations are individual tests, not independent validation of every vendor claim. The NVIDIA investigation remains open in the cited statements, and no finding of a security breach is presented here.