AI News: Haiku-5.5, Visual ChatGPT, Grok’s New Trick, and more!

Summary
In this weekly AI news roundup, tech YouTuber Matt Wolfe recaps major announcements and model releases from the first week of October 2026 across Anthropic, OpenAI, xAI, Google, Meta, and open-weight developers. The video highlights Anthropic’s Claude Haiku 5.5 release, OpenAI’s rollout of GPT-6 with visual “Intelligent UI” and its daily ship sprint, Grok Bot’s multi-model backend and X monitoring features, and a rapid-fire rundown of new open models and developer tools.

What is shown

• Claude Haiku 5.5 Overview & Benchmarks [00:11]: Anthropic’s announcement page, benchmark tables comparing Haiku 5.5 against Haiku 4.5, GPT-6 Luna, and Sonnet 5.5, and Artificial Analysis charts showing cost and intelligence rankings.
• BuseyBench Tests [03:24, 25:15]: SVG Gary Busey rendering benchmark runs for Claude Haiku 5.5 (score 4.6, 42s, $0.004) and Mistral Large 4 (score 5.8, ~2.5 min, $0.016).
• Claude Web Interface [03:38]: The model selector menu showing Claude 5.1/5.5 options (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5) and effort levels (Low, Medium, High, Extra).
• Claude Dashboards & Motion [04:29]: Anthropic’s announcement and video demos showing live BI dashboards connecting to BigQuery/Snowflake/Salesforce, along with automated video explainer generations.
• Hyperagent Multi-Agent Workflow [07:24]: Wolfe demonstrates creating three customized agents (Scout for research, Echo for interview strategy, and Vox for production rundown) to prepare for a creator interview, showing agent execution and the multi-agent “Rooms” workspace.
• OpenAI 28-Day Ship Sprint & GPT-6 Intelligent UI [10:02, 10:23]: Thibault Sottiaux’s X post announcements; official video demonstrations of ChatGPT’s “Intelligent UI” creating interactive bicycle breakdowns, recipe meal cards, vacation wardrobes with packing checklists, and bill-splitting widgets.
• ChatGPT Work & Permissions [14:02, 18:14]: Demonstrations of ChatGPT Work’s permission settings (“Approve for me” auto-review) and live steering mid-generation when changing an image prompt from a dog to a wolf.
• Meetings Plugin & Decisions API [14:47, 15:46]: ChatGPT’s meeting transcription plugin interface; OpenAI’s Decisions API interface displaying structured outputs (predicates, choices, scores) powered by GPT-6 Luna.
• Grok Bot & Multi-Model Integration [20:18, 21:44]: Elon Musk’s X post detailing third-party model support (Opus 5.5, Midjourney, Suno); Grok Bot’s native X search and monitoring analysis; and the Chief of Staff multi-bot interface delegating tasks to X-Bot, Research Bot, and Email Triage.
• Rapid Fire Launches [24:37]:
  • Mistral Large 4 (“Le Chonk”) 1T open-weight model announcement and DeepSWE benchmark table [24:37].
  • Reflection AI’s 501B open-weight model “Beam” announcement and reasoning benchmarks [25:27].
  • Google Playground game creation platform, showing the prompt UI and a playable demo of Skyfall Defender [26:39].
  • Google AI Edge Foresight local on-device meeting transcription application for macOS [27:29].
  • Google Cloud Gemini Agent announcement [28:23].
  • Google SynthID Detector web portal for identifying AI-generated media [29:35].
  • Hark Pro personal desktop agent demo by Brett Adcock [29:54].
  • Meta open-sourcing code for Muse AI hardware gadgets [31:23].
  • Anthropic’s updated Usage Policy prohibiting sustained cruelty toward Claude models [31:40].
  • Nobel laureate Daron Acemoglu’s article projecting AI will replace only 5% of human work in 10 years [33:33].

Claims & numbers

• Claude Haiku 5.5: The presenter notes it costs $0.10 per million input tokens (up to 100k tokens, then $0.50/M) and $0.50 per million output tokens (up to 100k tokens, then $2.50/M), roughly a 90% price cut compared to Haiku 4.5 and about a quarter the cost of Sonnet 5.5 [01:41].
• Claude Haiku 5.5 Performance: Scores 1620 on GDPval-AA v2.1, 1578 on AA-Briefcase v1.1, 72.4% on OSWorld 2.1 offline subset, and 57.4% on Humanity’s Last Exam (with tools) [00:39]. Artificial Analysis rates its cost per task at approximately $0.21 [02:54].
• OpenAI GPT-6 Speed & Pricing: Day 1 optimizations increased default inference speed by ~50% across GPT-6 Astra and GPT-6.1 Sol [13:30]. GPT-6.1 Sol Ultrafast runs up to 8x faster than standard Sol at $12/M input and $60/M output tokens [17:16].
• OpenAI API Rate Limits: Streamlined from five tiers into three: Build, Launch, and Grow [16:53].
• OpenAI Decisions API: Powered by GPT-6 Luna, offering decisions up to 10x faster than standard API responses [15:47].
• Mistral Large 4: Contains 1 trillion total parameters with 52 billion active parameters and a 1M token context window [24:47].
• Reflection Beam: A 501 billion parameter sparse MoE model with 23 billion active parameters [25:34].
• Economic projection: Cites Nobel laureate Daron Acemoglu’s estimate that only 5% of human tasks will be automated by AI over the next 10 years [33:46].

Notable quotes

• [02:05] “So if you need fast and cheap, but not necessarily the most intelligent model ever, well, Claude Haiku 5.5 sort of falls in line with, you know, like a Luna model.”
• [10:05] “Over the next 28 days, each day we’ll either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset.” (Quoting Thibault Sottiaux)
• [32:00] “So you’re not allowed to, like, be mean to the models anymore with Anthropic? I don’t know, I just thought that was really interesting.”

Assessment
This is a standard weekly news analysis and software review video by an independent creator, featuring a mix of live interface walkthroughs, slide reviews, and an integrated sponsor demo. Wolfe demonstrates live features in the browser (such as Claude’s selector, ChatGPT Work steering, and Google Playground), while relying on published blog posts and vendor demos for features that are enterprise-restricted or unavailable on his consumer plans.

Described by gemini-3.8-flash on 2026-10-10 from the video’s audio and frames.