Summary
Matt Wolfe's weekly 31-minute roundup, recorded on Thursday 15 Oct 2026. It is built around his own hands-on tests rather than straight reporting. He opens with Krea 3, an image editor he calls the most fun thing he used all week. Then he goes through a run of new language models, scoring each with his informal 'Busey Bench' (asking for a picture of Gary Busey, judged by an LLM) and a one-shot 'Mega Bonk' game test, and finishes with a rapid-fire round of smaller launches. The benchmark and price figures come from the companies or from the Artificial Analysis aggregate index. His own tests are anecdotal and rate looks and style as much as capability. One segment is a paid sponsorship for Genspark Design, and he says so on screen.
- Krea 3 (public beta, krea.ai) lets you select a region of an image and describe a change: recolouring a phone, reddening eyes, swapping a logo, rewriting text. It returns three variations in about 10 seconds and leaves the rest of the image alone. Wolfe's view is that it edits more cleanly than ChatGPT Image or Nano Banana, and that the others will probably catch up. This is his judgement from one session, not a measured comparison.
- Two open-weight image models also launched. Black Forest Labs' Flux 3 (transcribed as 'image clin') is claimed to run on a consumer RTX 4070, and on a DGX Spark to make an image in 7.3 seconds or edit one in 16.2. Alibaba's Qwen Image 2.1 Turbo is a faster checkpoint that trades some quality for speed. Both can be tried free on Hugging Face. Wolfe calls image generation 'fairly solved' and says editing is now where the gains are.
- Google Gemini 4 Argon (announced 30 Sep 2026) is reported to have caught up with GPT-6 Astra and Claude Opus 5.5 on most benchmarks. API pricing is a promotional $2 per million input tokens and $10 per million output until the end of November, after which he says it roughly doubles. On Artificial Analysis it scores 53, five points below Opus 5.5, but at a much lower cost per task. His Busey Bench run cost about 34 cents and placed second on his leaderboard. Google's team had told him it was good at 3D-printable models, and it produced a printable STL bust of Gary Busey from a single front-facing photo.
- Meta Muse Wonder: Meta reports 79% on a coding benchmark Wolfe calls the one most people watch (rendered as 'Deep Suite' in the transcript), which would be the top score, plus state-of-the-art results on several legal and finance benchmarks. Pricing is $4 per million input and $20 per million output, about the same as Opus 5.5. Artificial Analysis puts it level with Claude Sonnet 5.5, two points below Opus 5.5, and above Gemini 4 Argon, GPT-6 Astra and Fable 5.1. He cites the developer Theo ranking it 'S tier' alongside Opus 5.5, but says he has found it only 'kind of okay'. It came 14th on his Busey Bench, and its Mega Bonk game worked but looked plain.
- Four open-weight language models arrived in one week. Meta's Muse Spark Open is an 816-billion-parameter model with a 1-million-token context window, optimised with Apple to run on an M5 Ultra Mac Studio (he estimates about $10,000 of hardware), and billed by Meta as the most capable US open-weight model. Microsoft's MAI Base 2 (80B active parameters) has finished pre-training only, so it is not available or fully benchmarked. Nvidia's Nemotron 3.5 Super VL (120B) scores 26 on Artificial Analysis and sells itself on speed. Paper Instruments' Ultramarine One Flash claims 'Fable class' knowledge work at 60 times less than Fable, on benchmarks Wolfe calls cherry-picked, priced at $0.15 per million input and $0.50 per million output.
- OpenAI's '28 days of shipping' (announced 4 Oct 2026: one improvement or a usage reset each day) had a weak second week, in his verdict. The week brought composer predictions (it pre-types your next prompt), creating a 'dot' from your phone, Codex handing work to your dot, a larger GPT-6.1 Soul context in Codex/Work (500,000 tokens, up from 272,000), a claimed 60% cut in desktop storage, and on day 11 extra compute to reduce 'model at capacity' errors. Several days were resets only. He reports users complaining that bug fixes and basic reliability were being passed off as features.
- Decision models are a new category: models that return calibrated probabilities over fixed answers (yes/no, multiple choice, ratings, rubric grading of AI output) instead of generating text. After Jev and OpenAI's own, Microsoft released Decision 1 and Bespoke Labs released Nimble V4 and V4 Max. Both claim to beat Jev, but Wolfe says they all do essentially the same job.
- In the rapid-fire round: SpaceX AI's Grok Voice 3.0, which he used to build a working phone-support agent for Future Tools in its sandbox, and whose main selling point is how quickly it answers; ElevenLabs Instant Soundtracks, which watches an uploaded video and scores it; Facebook letting users type in interests to tune their feed; a Google Circle to Search check for scam texts on Android; Kylon, a chat-style tool where several agents and people work together (he had a demo); Apple's new Apple TV 4K and a counter-top 'Home View' screen with Siri, neither of which he has tried.
- AI and gaming: Microsoft has started Xbox Labs to experiment with AI game mash-ups and generation, and Epic's Tim Sweeney says the company has a 'huge pivot underway' to retool for 'maximum AI acceleration' without cutting staff. Wolfe notes that gamers are strongly hostile, and predicts audiences will accept AI-written code sooner than AI-made art, sound and voices.
Why it matters
The frontier is crowding at the top. In one week Google and Meta both arrived at roughly the level of Opus 5.5 and GPT-6 Astra, and the gap between them is now more about cost per task than raw benchmark scores, which supports the Handbook's existing principle that leadership rotates and should be judged task by task. The open-weight tier has also moved up to frontier-sized models that run on a single high-end desktop. For content creation, the news is a shift from generating images to editing them precisely, where Krea 3 currently leads in one tester's hands. Treat every benchmark here as vendor-reported or aggregate, and the tests as one creator's impressions.
KGcuuhqmPhI-transcript.txt