Seedance 2.5 Inches Closer

Hey gang, welcome back to Theoretically News!

Well, it's been quite a busy week. We had a pretty surprising drop from Meta (more on that down in the channel section), and while Seedance 2.5 hasn't quite been released yet, there's some new movement on that front. But really, the week belongs to the LLM boxing match: OpenAI's GPT-5.6 vs Claude Fable!

Pretty packed week. Check it out!

Seedance 2.5 inches closer and closer, with some interesting images dropping today showcasing the model being used on Jimeng — the Chinese version of Dreamina.

While this doesn't look like a full rollout on the Jimeng site, it is possible that some users are getting very early access to the model — from most reports, it's not in wide rollout yet. Interestingly, it looked like the only available resolutions were 480 and 720.

So if you're experiencing some FOMO, at least you know you're experiencing lower-resolution FOMO.

Advance rollouts of the Seed models in China are fairly common, but Western releases generally only lag by a few days. So I'd suspect 2.5 will be available for all sometime next week.

That said, again, this looks to be a SMALL group of users:

There were rumors that the model was set to release this week but was delayed by ByteDance for unspecified reasons. To be honest, I wouldn't read too much into it — from my own experience in early-access groups for new models, it always feels like everyone is flying by the seat of their pants. No matter how big or small the model.

In other interesting news, there was what appears to be a leaked promotional video for the 2.5 release, showcasing a slider that can be pulled out to 180 seconds for a single output.

In my previous Seedance 2.5 coverage, there was some speculation that this would be a beta feature — and sure enough, over on the official Dreamina page (which still looks to be a bit under construction), the long-video mode is explicitly labeled beta.

In terms of pricing — at least from what we could see — a 30-second 2.5 generation ran about double a 15-second 2.0 generation, which more or less tracks.

That said, pricing around Seedance can vary wildly — API, main platform, third-party platforms — so personally, I'm just going to wait until release to start complaining about the price. Which we all will, because it will certainly not be cheaper than 2.0.

All that said: I'll place my bet on next week being Seedance 2.5 week.

Not to say that ByteDance didn't release anything this week: Seedream 5.0 Pro landed Wednesday. I haven't gotten the chance to play around with it very much yet — but it's interesting that while ByteDance continues to dominate the video arena, and the recently released Seed Audio 1.0 is kind of killing it on the audio side, the Seedream image line hasn't quite made a major impact on GPT Image 2 or Nano Banana 2.

That's not to say Seedream 5.0 Pro is bad. It actually ships with a number of very interesting features, including:

Although at this time, the Layers functionality hasn't actually been activated — at least not on ByteDance's own platform, where launch-day testers couldn't view or export the separated layers.

The overall sentiment seems to be that it's... okay. While not bad, it's still edged out by GPT Image 2 and Nano Banana 2. That said, once the Layers feature is actually switched on, that may open up a powerful use case as an editor.

Kind of like Google's Omni — which I don't particularly feel is a strong generator, but does make for a very good video compositing editor.

I'll try to spend a little more time with Seedream 5.0 Pro over the weekend to see where it's really shining.

On the LLM side, this was a very big week, between the release of OpenAI's GPT-5.6 — and in particular its Sol tier, which punches at a weight just below the now-legendary Claude Fable.

I'm not going to go full benchmark-bro on this. But I will say that playing around with GPT-5.6 Sol — and in particular the new unified ChatGPT desktop app, which folds Chat, Codex, and the new ChatGPT Work agent into a single app — I'm finding it to be a pretty good model.

Anthropic's answer to all of this was to extend the Fable trial period before the model gets locked out of standard plan usage. So at least until the 12th of the month, we have access to two of the most powerful models on the planet. And that's a lot of fun — especially when you make them argue.

What happens after the 12th? We'll see — but I'm hopeful that Anthropic extends Fable usage a little bit longer, or at least offers a reasonably priced plan that won't send you into Chapter 11 for trying to vibe-code a fitness tracker app you will never actually use.

In the meantime, I’m working overtime building as much as I can!

Waiting in the wings: Meta released Muse Spark 1.1, and xAI released Grok 4.5. While neither model competes at the level of 5.6 or Fable, they also aren't very far behind.

This week's Cool Thing is The Visual History of AI Film & Video — a curated, interactive timeline of this entire weird medium, built by Rich Klein (@RichKleinAI). It runs from 2019 ("Before the Dawn") through 2026 ("The Long Tail"): roughly 120 entries and 80-plus embedded clips, from Will Smith eating spaghetti, through Sora's launch and app shutdown, all the way up to the Seedance era — model releases, festivals, viral moments, and the films in between.

Why It's Cool: we all lived through this timeline in weekly increments, and scrubbing through it in one sitting is genuinely vertigo-inducing — two-second blurry nightmares to Lincoln Center premieres in about three years. Klein is upfront that it's a personal take rather than an institutional archive, and honestly, that makes it a better read.

And a big thanks to Rich for popping Dragon Blue into the Timeline— Flamethrower Girl is now officially marked in the history books!

Only one video this week, but it was a big one!
Meta's FREE "Banana Killer" & My AI Video Tool (Also Free!)

First half: Meta actually shipped something! Muse Image is live, free, at meta.ai — and I ran it through the full gauntlet: prompt adherence, text rendering, image editing, the Elder Scrolls 6 imagination test, and its visible "thinking" mode (including watching it talk itself out of a correct answer).
Spoiler: the Banana survives — GPT Image 2 and Nano Banana 2 still hold the top spots — but Meta's model is free, real, and closer than expected. We also break down the first Muse Video samples, since that one is still announcement-only.

Second half: I built you a free tool!

It turns any footage into an OpenPose skeleton + depth map, which gives you real motion control over your AI video generations — multi-character tracking, in/out points, and what I'd call the hardest video-to-video shot I've ever attempted on the channel. Works with Seedance, Runway, and anywhere else that takes a video reference. It's on Gumroad — put $0 in the box, it's yours, source code included: grab it here

It isn’t PERFECT— but I think it’s off to a good start, and since the code is all there, you can continue to play with it and build out your own version if you’d like!

That's the week!! NEXT week I think is going to be a bit of a monster— Send Coffee! I’ll be needing it!!