Opus 5.5 vs GPT-6 Sol, Meta's Muse Charm, and a PixVerse R2 World You Can Talk To
Opus 5.5 and GPT-6 Sol dropped on the same day, Meta turned Muse into hardware with the Muse Charm and a $1,299 Meta VR Glasses, and PixVerse R2 lets you prompt inside a live world. On the creative side: Alibaba's unnamed November video model (the "Seedance killer" label isn't Alibaba's), a reported Kling update, Seed Audio 1.5, Fish Audio's Drama 3 preview, and Jordan Rudess's jam_bot.
Anthropic and OpenAI shipped Opus 5.5 and GPT-6 Sol (plus Luna) on the same day, Meta Connect turned Muse into a hardware story, and our side of the fence had a lot more “coming soon” than “try it now.” The full rundown’s in the video; here’s the week in order.
Two hedges up front: Alibaba’s November video model is unnamed, and “Seedance killer” is the internet’s label, not Alibaba’s. And the Kling item is a report I can’t confirm.
Opus 5.5 and GPT-6 Sol land the same day
Setting benchmark charts aside, because honestly, who believes benchmarks anymore, the vibe is that Anthropic is so back. Opus 5.5 came out Sept 22, about 90 minutes ahead of OpenAI’s GPT-6 Sol and Luna, per TechCrunch. The pitch is that it competes with Fable on intelligence at a fraction of the cost.
The other big win, going by the general read, is that 5.5 seems to have removed the annoying personality that’s plagued the Opus models since 4.6 was superseded. I tend to agree: I’ve been playing with it a lot, and it’s just a lot nicer to work with.
Sol comes in significantly cheaper at $2/$10, half Opus 5.5’s per-token price. The sentiment is that it taps out of projects a little quicker, but it’s cheap and it ships. The complaint is that it’s almost too concise, to the point of bland. I don’t necessarily agree, and you can tune that in the settings. That may be because I’ve been leaning a lot more on voice in ChatGPT and Codex, where Anthropic is sorely behind.
ChatGPT Voice gets your apps, and Siri gets a one-line review
OpenAI also finally connected Voice to mail, calendar, and Slack, which always felt like a bit of an oversight. It brings us closer to an assistant that’s actually helpful day to day. At least more so than the new Siri: remember Alexa’s big update, and then it was still pretty much Alexa? That’s my mini review.
Meta Connect: Muse, the VR glasses, and the Muse Charm
Muse, Meta’s assistant, came storming into the hardware game. It’s been a blockbuster by Meta’s standards, overtaking the ChatGPT app on the US App Store’s free chart. (I looked at the family’s image side in July with Muse Image.)
| Product | What it is | Price and timing |
|---|---|---|
| Meta VR Glasses | Glasses plus a roughly 300g puck on a cable | $1,299; ships spring 2027, no pre-orders |
| Muse Charm | An AI Tamagotchi with real-time voice and its own 5G | No price yet; ships December |
To me the VR glasses look like what the Apple Vision Pro should have been all along, at a sane $1,299, though you can’t buy them until spring 2027. The Charm is what caught the news cycle: basically Muse without a phone or a computer. AP says it works without a phone; Bloomberg says it syncs through the Muse app. (Meta’s Connect recap has the rest.) Meta does seem to have won the week with the AI Tamagotchi.
Dev Day next week, and the race to be your AI best friend
OpenAI has Dev Day next Tuesday, Sept 29. The rumor, and it is a rumor, is that we’ll see something called Aeon, OpenAI’s answer to Muse and Grok Bot, though The Information says a Muse-style assistant is only at the discussion stage. I’m curious how OpenAI differentiates itself. Any way you slice it, the next battleground is to become your new AI best friend.
Alibaba’s video model is coming in November
At its Apsara conference, Alibaba showed a new video model it says is coming in November. To me it looks like it’s in around Seedance quality, but all we have to go on is bootleg phone footage from the event. Alibaba made no Seedance comparison.
It appears to be the successor to Wan. Coverage pairs it with Qwen 4, but that’s reportedly still in training and this model has no name. It’s the standard new-video-model bingo card:
| Alibaba said | Alibaba didn’t say |
|---|---|
| Longer duration, with scene, character, and visual consistency | A duration number |
| Better control, more complete narratives, smarter reasoning | Pricing, open weights, a name |
They also predicted that within three years a native unified omni-modal model will emerge: no Seedance, no MiniMax H3, just one model. I’m presuming Qwen.
Kling is back on the menu, reportedly
Per screenshots I can’t confirm or deny, Kling has an updated Omni model coming, with generations up to 120 seconds at 720p and 1080p. “4.0” is the name attached to it, and I’d keep a question mark on that. For context, current Kling models top out at 15 seconds. What’s live is on the Kling page.
Nano Banana 2.5 Flash shows up in Google Flow
Via TestingCatalog, traces of Nano Banana 2.5 Flash have turned up in Google Flow. Traces, not a launch, and I don’t have access yet. It’s not Nano Banana 2 Pro and it’s not Nano Banana 3, and Google really seems stuck in Flash land since Google I/O. I’m hoping Gemini 4 drops soon, then a flood of big updates. Current state: Nano Banana.
Odyssey Agora-2: Diablo II with AI agents
Agora-2, from Odyssey, is live now: AI agents playing alongside you and other humans, up to 20 humans and agents in one world per Odyssey, and yes, everyone is playing Diablo II. The obvious read is “great, an AI NPC,” but Odyssey is after a sandbox for how AI systems interact with each other and with us. I’m not sure Diablo is the best place to teach an AI our behavior: when threatened, humans stab things, and when feeling low, they chug colorful drinks.
PixVerse R2: a world you can talk to
PixVerse released R2, and I haven’t had much time with it. It looks like a pretty big step up in visual quality. The bigger update: aside from the Genie 3-esque controls, you can actively prompt within the generated world, which opens the door to chatting with a character in it. Alex Prompter captured a good example, though it looks sped up a little, and when I played with R2 it felt a bit sluggish. PixVerse published no resolution or frame-rate numbers.
We’re still early, but I think we’re starting to get to a breaking point, where this genuinely resembles a holodeck, minus the holograms. It’s a massive step up from the wonk we got in real time back in January, when things randomly caught on fire and coherence was, let’s call it, a very surreal suggestion. More on the field: Lucy 2.5 and MiniMax H3 Max.
Seed Audio 1.5 is landing on Pippit
An update is coming to ByteDance’s Seed Audio, which I covered a few videos back in the audio-first SeedAudio workflow and found very good. Think Seedance, but audio. 1.5 was apparently announced as 2.0 and then walked back half a point, and it’s landing on Pippit, which I hadn’t realized was another ByteDance platform. ByteDance, why do you have to make things so confusing? Audio isn’t Seedance’s strong point, and 1.0 really does seem to help there.
Fish Audio Drama 3 (preview): usable today
Fish Audio released Drama 3 in preview, calling it the most controllable text-to-speech model ever. I’ve been a big fan of Fish’s emotive readings, and the sample sounds pretty good overall. It’s in the model dropdown now. When I logged in it was there, free plan included, though I don’t know how many generations that gets you.
Jordan Rudess and jam_bot
Rounding out with a project I’ve tracked for a while: my pal Jordan Rudess, yes, that Jordan Rudess, the Dream Theater keyboardist who can play guitar and keyboard at the same time. He’s spent the last few years with the MIT Media Lab on jam_bot, or as I like to call it, Jordan bot, an AI improvisation tool trained on his playing.
It isn’t a diffusion model trained on recordings, the way Suno works. It’s a model trained to think like Jordan, generating music in real time from the notes and phrasings he’d most likely choose. It’s research, not a product, though I’d love to make it suffer through me playing “Dance of Eternity” at 70 BPM.
The concert is Saturday, October 3 at MIT, with Jordan, jam_bot, live musicians, and the MIT Chamber Chorus (tickets). I appreciate that Jordan isn’t running from the technology; he’s figuring out how to put it into performance. If you go, don’t be the person screaming “Freebird.”