Matt Wolfe’s 33-minute roundup combines model comparisons, an original website-design test, and a rapid tour of new AI products.

Cheaper models, uneven creative results

  • Wolfe presents Anthropic’s IPO and staged insider-share unlocks as a test of market confidence; that interpretation remains speculative.
  • GPT-6.2 Astra and Sol emphasize lower task cost, while Google positions Gemini 4 Neon between its largest and fastest models.
  • OpenAI lists Astra at $8/$40 and Sol at $2/$10 per million input/output tokens; Google lists Neon at $3/$15.
  • Wolfe likes Astra’s game-demo visuals, while Neon’s demo has working movement and less impressive graphics in his assessment.

An open-ended website prompt exposes repetition

  • He asks Astra, Sol, and Neon to build an impressive website with the subject left to the model.
  • Astra and Sol produce similar deep-sea exploration sites, including overlapping language and limited interaction.
  • Neon produces a watch-themed site with a hover effect and magnifying-glass interaction, which Wolfe prefers.
  • These are creator-run examples and aesthetic judgments, with no repeated blind evaluation shown.

Microsoft and source-based websites

  • Microsoft’s MAI Code 2 targets everyday coding; its relative performance and cost figures come from Microsoft’s selected evaluations.
  • MAI Image 2.7 emphasizes image editing; MAI Realtime 1 starts at one cent per audio minute, with reasoning billed separately.
  • Gemini Notebook turns imported sources into an interactive site; Wolfe demonstrates a newspaper-style page with charts and source filters.
  • Hyperagent’s iMessage workflow appears in a clearly identified sponsored segment.

Video, music, and desktop agents

  • HeyGen Avatar 6 advertises long talking-avatar videos from a photo; its “84.4%” figure is a transformed comparison metric, not 84.4% fooled viewers.
  • Suno Visuals generates music-linked images and clips up to 15 seconds; visual credits are purchased separately.
  • Gemini’s Mac computer use adds permissioned background interaction with apps; Wolfe has not tested it in this episode.
  • Wolfe also reports expanded Claude Duet access and custom themes for local Claude Code desktop sessions.

Actions still need boundaries

  • Google’s waitlisted Stitch Loop proposes improvements for security, latency, and accessibility through reviewable changes.
  • ChatGPT Wallet uses user-approved payments and substitute card details; the current launch differs from future fully delegated purchasing.
  • Live dictation exposes text while speaking, and ChatGPT for Sales bundles sales workflows with partner integrations.
  • Genie 4 adds prompt-driven world changes and generated audio; the separate Genie 3 update extends sessions to three minutes.
  • Caveat: Several segments summarize vendor announcements. The generated news site also shows a valuation headline inconsistent with Wolfe’s earlier narration.

“Use your eyes. Don’t pay attention to these numbers.” — Matt Wolfe, 14:36–14:41, on judging visual examples