As of: 2026-10-10 14:45 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code). Human editor: Adam Bicz.
Canonical page: https://postcutoff.com/v/worldofai-gpt-7-bel-mythos-5-1-mistral-large-4-news/

# GPT-7 'Bel' First Preview, Mythos 5.1 Access, Mistral Large 4, Claude Code Update, & More! AI NEWS

WorldofAI, 7 October 2026, YouTube. 211,103 views as of 10 October 2026.
Kind: Review.
Watch: https://www.youtube.com/watch?v=-mv1Tf26Vms

## Description (written by Gemini from the video)

**Summary**  
In this news roundup, the presenter (World of AI) covers major AI developments from early October 2026, focusing on OpenAI's release of 722 AI-generated research math papers from an internal frontier model and Day 2 of OpenAI's 28-day shipping sprint. He also analyzes Mistral AI's 1-trillion-parameter open-weights model Mistral Large 4 ("le Chonk"), Anthropic's Cyber Verification Program granting access to Claude Mythos 5.1, Claude integration with Google Workspace, and updates from Google, Zapier, and independent developers.

**What is shown**  
- **OpenAI Math Research Release** [00:20, 06:15]: The `openai/math` GitHub repository containing 722 mathematical preprints and Lean formalizations across 372 result families; animations explaining claimed results including the free group factor problem [07:11] and the plane coloring problem [08:11].  
- **OpenAI 28 Days of Shipping (Day 2)** [01:13, 09:46]: Announcements from Thibault Sottiaux regarding free Auto-review, consolidated API tiers (Build, Launch, Grow), the Meetings plugin for ChatGPT macOS desktop, and the real-time Decisions API [11:39].  
- **Mistral Large 4 ("le Chonk")** [01:42, 13:00]: Model launch details and benchmarks, including 3D WebGL scene generation tests ("Paris at Night", "The Aquarium", "The Chonk") [14:40, 16:40], a 19-challenge CTF speedrun UI [14:36], and automated malware triage analyzing a Cobalt Strike loader [15:14].  
- **Google Nano Banana 2.1 & EmbeddingGemma 2** [02:22, 18:50, 20:34]: Image comparisons between Nano Banana 2.1 and GPT Image 2.5 across five prompts; on-device multimodal search demo testing EmbeddingGemma 2 on an Nvidia DGX Spark indexing 300 animal images [21:11].  
- **Anthropic Claude Mythos 5.1 & Workspace** [04:51, 22:01]: The Cyber Verification Portal UI showing eligibility criteria and tiers (Defense, Red Team, Specialized Access); sidebar integration of Claude inside Google Docs, Sheets, and Slides.  
- **Claude Code & ArtCraft** [22:45, 23:32]: Setting subagent reasoning effort in Claude Code terminal; GitHub repository and UI of PhotoCraft, an open-source Rust recreation of Photoshop built using Claude Opus 5.5.

**Claims & numbers**  
- The presenter notes that OpenAI published 722 AI-generated mathematical manuscripts across 372 result families after evaluating an unreleased model on approximately 4,000 open research problems [00:26, 06:26].  
- The presenter states that the average successful OpenAI math result required roughly 3 hours of ChatGPT Pro–equivalent thinking compute [06:37].  
- According to OpenAI's Day 2 updates, Auto-review previously accounted for 2% to 10% of user plan limits but is now free, and the Grow tier qualification threshold dropped from $1,000 to $500 in monthly API spend [10:40, 10:56].  
- The presenter states that OpenAI's Decisions API operates up to 10 times faster than GPT-6 Luna via the response API [11:50].  
- Mistral Large 4 is presented as a 1-trillion parameter MoE model with 49 billion active parameters, a 1-million-token context window, and 256k output tokens, with open weights scheduled for the end of October 2026 [13:08, 16:21].  
- The presenter claims Mistral Large 4 generated three 3D benchmark scenes for $0.87 total over 66 minutes, compared to Claude Opus 5.5 costing $13.83 over 38 minutes [17:08, 17:55].  
- In a cybersecurity CTF speedrun, Mistral Large 4 reportedly solved 18 of 19 challenges and triaged a Cobalt Strike loader in 12 minutes [14:53, 15:50].  
- Nano Banana 2.1 reportedly generated benchmark images at approximately $0.044 per image in ~11 seconds, versus GPT Image 2.5 at $0.069 per image taking ~50 seconds [19:59].  
- EmbeddingGemma 2 is stated to be a 740-million parameter multimodal embedding model licensed under Apache 2.0 that achieved 82.5% precision (165/200 matches) on an animal retrieval test [20:42, 21:35].

**Notable quotes**  
- [00:07] "Starting off with OpenAI, we believe that we have gotten another glimpse at their unreleased model, potentially Belle, GPT-6.5, or GPT-7."  
- [04:54] "Anthropic just opened up access to one of its most restricted models. Mythos 5.1 is reportedly using the same underlying weights as Fable 5.1, but with significantly looser cybersecurity and biological safety restrictions."  
- [08:20] "Just three years later, we're discussing AI potentially producing mathematics at the frontier of human knowledge."

**Assessment**  
This is a third-party AI news summary combining official lab blog posts, GitHub repositories, and community benchmark runs into an editorial recap. The demonstrations rely on public announcements, third-party benchmark captures, and pre-recorded UI workflows rather than live interactive prompting during the video.

_Described by gemini-3.8-flash on 2026-10-10 from the video's audio and frames._

## Related

- 2026-10-06: [Mistral releases Mistral Large 4 ("le Chonk")](https://postcutoff.com/e/2026-10-06-mistral-large-4/)

## People in it

- [Thibault Sottiaux](https://postcutoff.com/person/thibault-sottiaux/), Leads core product and platform, OpenAI (previously led Codex)