Is This The Great Pause? Not For AI Video!

Hey gang, welcome back to Theoretically News!

Kinda bonkers week for AI Summer— with some highs and some lows.
In the span of about a week, OpenAI's next model got tucked behind a government access list— BUT: We just got some flashy specs! Google pushed its frontier model to July, and meanwhile — ByteDance casually flexed Seedance 2.5.

Plus: Google's Omni Flash pokes its head out, Krea 2 goes open source, and ByteDance's Seed Audio lands with a pretty interesting use case!

OpenAI previewed GPT-5.6 today — three models, Sol (the flagship), Terra (the balanced, cheaper tier), and Luna (the fast one) — and almost nobody can actually use them yet. It's a limited preview: OpenAI says it's starting with a "small group of trusted partners," which Axios pegs at around 20 organizations, reachable for now only through the API and Codex (broader ChatGPT and API access, the company says, is still "coming weeks" out).

So yes — there's a real announcement, a real system card, and a real price sheet. But…we still can’t use it.

More than that, the 1.5M-token context and the alignment-fix rumors are still unconfirmed — the system card talks safety tiers and capabilities (it rates all three models "High" risk on both cybersecurity and bio/chem), not context windows.

What keeps making me laugh are the benchmarks that read: It’s good! Really Good! But not a good as Mythos! Mythos BAD and SCARY!

That said, we DID get some pricing: Sol lands at $5 in / $30 out per million tokens, actually cheaper than Claude Fable 5's $10/$50. But— still pretty costly.

Second, and quieter: Google delayed Gemini 3.5 Pro to July, reportedly to fold in more real-world feedback from early testers and from 3.5 Flash. (Google declined to comment, so this one's a credible scoop rather than an official date — but the reporting is there)

My Take: A pause at the top can cut both ways. It's a window for Google to close the gap with the Gemini 3.5 Pro LLM— BUT also, this could be the stretch where the eastern models stop being "the cheaper option" and quietly become "the better option, full stop." Bytedance has clearly taken the lead in video with Seedance, so could LLMs be far behind?

Speaking of Google making a move: a preview of Gemini Omni Flash looks to be arriving in AI Studio soon — spotted in an announcement ahead of any formal rollout. The number getting passed around is "4K, 60 seconds," which... needs an asterisk.

Omni Flash is a real, already-demoed Google video model, and it does do 4K — but single-shot clips run a lot shorter than a minute; that 60-second figure comes from stitching generations together in Google Flow's timeline, not one continuous take. So: exciting, multimodal, genuinely 4K — just not a minute in one breath.

Why It Matters: It proves Google is still cooking. That said, the thing we’re all looking for is Gemini Omni PRO. Flash models are important, especially as frontier token costs rise…
But also…I know, it’s hard to get excited about Flash Models?

Some genuinely good news Krea open-sourced Krea 2 — currently the #1-ranked image model from any independent lab on Artificial Analysis.

It's a 12B model that runs on a 16GB GPU, and the install is genuinely painless now: it's a ComfyUI template you literally click — no dependency rabbit hole.

This is one of my favorite image models, so I'm thrilled to see it go Open Source. One thing to flag: the version you'll actually be running open source is the Turbo model — the fast, distilled one. (Krea also released a "Raw" checkpoint, but that one's really the base you fine-tune on, not the one you generate with.)

To make Turbo really sing, you're going to want to stack some LoRAs on top of it — out of the box it's good, but it's the LoRAs that take it where you want to go. I’ll try to get a tutorial on all that up soon!

The other half of ByteDance's big week: Seed Audio 1.0, an all-in-one audio model that generates voice, music, and foley/SFX in a single pass — zero-shot voice cloning from a short reference, multi-speaker dialogue with distinct voices, cross-lingual, up to about two minutes a go. It's exclusively on fal for now.

Full disclosure: I haven't put this one through its paces yet — though I was curious enough to jump into the thread and ask how it was being driven. TomLikesRobots ran a genuinely interesting test — using Seed Audio for dialogue and sound replacement on a Seedance output — and the results are worth a watch/listen. Give it a look here.

His early read: the voice acting and foley came out better than Seedance 2's native audio, and it's cheap enough to experiment with — a 15-second test ran him a couple of cents.

Why It Matters: My Audio gang knows: Sound is Half Your Picture!

Seedance Mania this week! Kicking off with the big announcement from the Volcano Engine Conference and covering all the details released about Seedance 2.5!
Plus: A look at Seedance 2.0 in 4k
AND— a new feature from the Kling gang that might be showing up in the next (hopefully big) update!

Check out the video here: https://youtu.be/s40M0IvBEpM

And we weren’t done with Seedance news yet…

Because there was also one BIG story from the conference that everyone seemed to have missed. And it’s a BIG one— one that might just reshape AI Video’s future. Really, I mean that. It’s big! No Spoilers if you haven’t seen it yet! Check it out here: https://youtu.be/1RB4snLk2LU

Well, that concludes Bytedance Week— the clear winners of the week, while everyone else seems to be reaching for the pause button.
Hopefully we’ll see some big summer energy start to whip up!

If you have any feedback or anything you’d like to see, please let me know: [email protected]

Or, just drop a comment on one of the videos! I honestly do see them all!