AI News: Opus 5.5, GPT-6 Sol, Jev, Muse and More!

Summary
Matt Wolfe presents a weekly AI news roundup recapping major industry announcements, including hardware and agent features from Meta Connect 2026, new frontier models from OpenAI (GPT-6 Sol and Luna) and Anthropic (Claude Opus 5.5), and SpaceXAI's Grok 4.7. He also analyzes TypeSafe AI's decision-focused "Jev" model, runs custom game-development and portrait benchmarks, and covers rapid-fire updates from YouTube, Microsoft, Google, and Spotify.

What is shown

• Meta Connect 2026 recap [00:26–08:41]: Meta's Muse agent (glasses integration, voice mode, Mac computer-use capabilities, connectors to Granola/GitHub/Box, custom email address, and the standalone "Muse Charm" teaser); Ray-Ban Meta Audio (camera-free glasses); Meta Ray-Ban Display features (turn-by-turn navigation, calendar, notification mirroring); and Meta VR Glasses hardware (100g, micro-OLED, external compute/battery puck, full holographic calling with legs).
• OpenAI GPT-Live-1 & Codex Demo [08:42–09:47]: Wolfe demonstrates an interactive voice-driven AI news application ("Airwave Live") in Codex using GPT-Live-1 and the Agents API, showing real-time background search and spoken interruptions.
• GPT-6 Sol & Luna Benchmarking [09:48–12:55]: Review of pricing and benchmarks on AutomationBench and Agent's Last Exam. Wolfe tests a 3D browser game prompt ("Megabonk 3D clone") in Codex using GPT-6 Sol Ultra (taking 24m 10s to output "Bonkbound") and evaluates SVG portrait generations on BuseyBench.
• Claude Opus 5.5 [12:56–18:14]: Benchmark charts (Terminal-Bench 4.0, Human's Last Exam, agentic coding); a 20-hour autonomous coding run generating a full 3D "Megabonk" game clone; JavaScript and Blender animations generated via Opus 5.5 (browser parsing by Addy Osmani, transformer visualization by Chubby, architectural model by Techartist, Large Hadron Collider simulation by Alexey Fateev).
• SpaceXAI Grok 4.7 [18:15–20:42]: CursorBench and cost evaluation charts; a simplified cube-and-pill Megabonk game test; BuseyBench portrait results.
• TypeSafe AI Jev [20:43–24:45]: Demonstration of "System One" decision-making returning type-safe structured values (choice, score, noul) rather than freeform text; a speed/cost test versus GPT-5.6 Terra (0.11s vs 8.56s); community demos including smart drag-and-drop file organization, inbox urgency scoring, live comment moderation, and real-time video game input.
• Rapid Fire Section [26:48–33:04]: "Made on YouTube" AI features (custom feeds, Ask YouTube search, live auto-dubbing, Ask Music, YouTube Studio thumbnail/editing tools); Microsoft Copilot modes (Home, Code, Autopilot); Google Gemini 3.8 Live Avatars, Gemini 3.8 Flash TTS with Voice Design, Gemini Omni in Google Vids, and Project Suncatcher (deploying TPUs to orbit on SpaceX rockets); Spotify Taste Profile.

Claims & numbers

• Meta VR Glasses: The presenter states they weigh 100 grams, feature a 5K Infinite Display built on micro-OLED with 37 pixels per degree, and are scheduled to go on sale in Spring 2027 for $1,299.99 [05:41, 07:04].
• GPT-6 Sol & Luna Pricing: The presenter states GPT-6 Sol API pricing is $2.00 input / $10.00 output per million tokens (50% cheaper than GPT-5.6 Sol at $4/$20), and GPT-6 Luna is $0.10 input / $0.50 output per million tokens (cut from $0.20/$1.20) [10:13–10:33].
• Claude Opus 5.5 Pricing: The presenter states base token costs are $4.00 input / $20.00 output per million tokens (down from $5/$25 on Opus 5 and $10/$50 on Fable 5.1), with fast mode priced at $8/$40 [13:51–14:26].
• Grok 4.7 Pricing: Benchmarks show API pricing at $2.00 input and $6.00 output per million tokens [18:42, 18:57].
• Artificial Analysis Rankings: On the Intelligence Index, Opus 5.5 Max ranks #1 (58 points), above Fable 5.1 (55), GPT-6 Astra (53), GPT-6 Sol (48), and Grok 4.7 Extra High (46) [19:46–20:09].
• Cost Per Task: The presenter displays Artificial Analysis metrics showing GPT-6 Sol at $1.06 per task, Grok 4.7 at $3.74 per task, and Opus 5.5 Max at $5.88 per task [21:03–21:24].
• TypeSafe AI Jev Metrics: The presenter shows Jev costs $0.042 per million input tokens, with output tokens listed as free, achieving end-to-end response times of 70 ms to 500 ms [23:44–23:54].

Notable quotes

• "This is officially the new state-of-the-art model. This is pretty much the best model there is, kind of hands down at the moment." [13:13]
• "It worked for almost 20 hours building, testing, building, testing, building, testing, and the result is pretty mind-blowing." [14:30]
• "Existing LLMs are optimized for human preference: write-ups and chat responses that human raters prefer. Now Jev, this new model, is calibrated for decisions." [22:56]

Assessment
This is a tech news recap and hands-on review video hosted by an independent creator, featuring a sponsored integration demonstrating OpenAI's API. Several live demonstrations (Codex game development, BuseyBench, Jev API races, and web-based games) reflect real model outputs, while external demos and Meta Connect clips rely directly on company promotional material and developer social media posts.

Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.