Kling 4 Early Look, Gemini 4 Argon, Ideogram 4.5, and Eleven v4
Kling 4 is in early access — Flash model only, 720p, Ultra annual subscribers first — and I ran it through the usual tests: lip-sync drift looks fixed, performances run a little flat, and it's maybe not the Seedance killer some had hoped for. Plus HeyGen's penny-a-second video model, Google's Gemini 4 Argon (announced, not released), Ideogram 4.5's precise editing, and Eleven v4.
Four is apparently the number of the day: Kling 4, Gemini 4, Ideogram 4.5 and Eleven v4 all turned up together, plus a penny-a-second video model from HeyGen. The lead is an early, tested look at Kling 4. It’s early access and the Flash model only, so read it as a first look, not a verdict on the full release. Gemini 4 is announced, not released, so nothing about it is tested. The full rundown’s in the video.
Kling 4: early access, Flash model only
It has been a minute. Kling’s last big version jump was 3.0 back in February; we got 3.0 Turbo in June, but as far as a whole new generation goes, it’s been a while. As of now we have V4. Well, kinda. Kling 4 has entered early access, and sadly this is the Flash model. The initial rollout is for Ultra subscribers on the annual plan, with the larger “Big Daddy” model planned for sometime in October. That’s not uncommon for Kling. (Side note, and a bit inside baseball: Kling is also working toward an IPO.) Output is 720p for now.
| Test | Mode | What I saw |
|---|---|---|
| Blue business suit, jaywalking ticket | Text-to-video | Pretty solid; some unprompted cop motion |
| FBI agent, Pacific Northwest diner (Twin Peaks dialogue) | Image-to-video | Dialogue handled with room to spare; invented the waitress |
| Flamethrower Girl’s birthday cake | Omni | Candle lit on entry, doesn’t melt; sink stays put |
| Tom Malloy’s office, noir | Omni | Eyelines match; invented a gray jacket; a little flat |
| Vampire hunter | Keyframes + Omni (up to ten references) | Holds together pretty well |
Text-to-video and image-to-video: the ticket and the diner
On the text-to-video side, our man in the blue business suit is about to get a jaywalking ticket. Not so bad. The police officer does a motion I definitely didn’t prompt for; my best guess is he’s trying to ID the guy from a photo. Still, pretty solid.
The diner test comes down to two things. Can it handle all of that dialogue? It did, with room to spare. And the waitress has her back to us in the initial image, so when she turns around the model essentially has to invent her. It did a pretty good job and held stylistic consistency. You’ll probably see some decoherence in these, but remember we’re looking at the fast model.
Omni: Flamethrower Girl’s birthday and Tom Malloy’s office
Kling’s Omni version of Flamethrower Girl has always been one of my favorites (the OmniDirector coverage has the history), so I gave her a role and brought her home in version four: a birthday cake and a flamethrower. Nailed it. The biggest issues: the candle is already lit when she walks in, and, I’m sure someone out there is upset, it doesn’t melt after she flamethrowers the entire tabletop. Pretty silly prompt. The upside is some improvement in spatial awareness: the kitchen sink is in the same place when she walks in and later in the generation.
Next, Tom Malloy’s office for the noir project. It’s simple in terms of camera setups, but it holds up spatially and the eyelines match. For whatever reason, Kling’s Omni model put Tom in a gray jacket that wasn’t provided or called for. Video quality, albeit only 720p right now, is pretty good. The performances, personally, I find a little bit flat.
The lip-sync fix, keyframes, and the early verdict
The big win: Kling 4.0 does seem to have solved the lip-sync drift problem that plagued 3.0.
On performance, which I still have to dig into, I think you can improve it by combining keyframes with an Omni reference, a method I’ve been using more and more. You get the best of both worlds, and for keyframe referencing, 4.0 allows up to ten references. The vampire-hunter scene is built that way, and aside from some edit-timing issues at the start of the clip, it holds together pretty well. My running theory is that giving it keyframes for your characters gives the model a better grip on what the performance should feel like.
So, an early look at Kling 4.0: maybe not the Seedance killer that some had hoped for. I’ll roll back with a deeper look once the Big Daddy model drops. Kling’s complete guide to 4.0 is worth a read in the meantime.
HeyGen Video: a penny a second, with an asterisk
HeyGen has released HeyGen Video, its first general-purpose video model, and yes, it’s another one built off MiniMax H3. We are definitely seeing a lot of those. The pitch is cheap and fast.
| Resolution | Launch price (through end of October) | Regular price |
|---|---|---|
| 480p | 1¢ per second | 2¢ per second |
| 768p | 1.5¢ per second | 3¢ per second |
That penny is a 50%-off launch sale, and it’s 480p only. I really can’t get mad. I haven’t tested it, but it’s an H3 fine-tune, so you can kind of figure where it’ll land. It’s available now through the API and ComfyUI, with HeyGen’s web interface coming later. HeyGen’s announcement also lists OpenRouter and Runware, and notes that reference-to-video costs double.
Gemini 4 Argon: announced, not released
The giant awakens: Google has returned with Gemini 4 Argon. Once again, this is on the announce side, not the widely released side. The benchmarks are impressive, but they’re Google’s own, and of course they put Gemini 4 ahead of OpenAI and Anthropic: per the announcement, ahead of GPT-6 Astra and Opus 5.5 on 12 of the 18 benchmarks Google reports.
One of Gemini’s secret powers in general is video understanding, and there Gemini 4 does look pretty impressive: a 91.7 on LVBench, the test for long-form video.
As for when it rolls out more broadly, we don’t know. Per Google, it goes to trusted cyber-defense partners first, then the paid API and AI Ultra, with no date. It’s no secret Google has been stuck in Flash land, so it’s nice to see a flagship model finally on the verge of being widely released.
To me, the bigger thing is what’s behind it. A big model probably creates momentum for the creative tools we actually use: Nano Banana 2 Pro, possibly a Genie update, maybe Google getting back into the video race with an Omni model, Veo, or whatever it ends up being called. Who knows, with Google’s naming. Those are hopes, not announcements. I’m also curious what Google has in store to match the new class of agentic assistants, like Muse and OpenAI’s dots.
Pour one out for the OG Nano Banana
Google has scheduled the shutdown of Gemini 2.5 Flash Image, our first taste of the banana. It goes away in the Gemini API on October 2, per Google’s deprecations page.
Ideogram 4.5: precise editing
Another quick hit on the image side: Ideogram 4.5 is out, and the focus is more precise editing. Ideogram’s pitch is editing the same image over and over without drift, with open weights promised. I haven’t had a ton of time with it yet, but Ideogram has always been a really solid model for me, so I wanted to try the editing.
I pulled up an image I generated in Ideogram about ten months ago, around the time of Gemini 3’s release, with all the garbled text you’d expect from ten-month-old AI image problems. I made some text-edit corrections and, because why not, swapped in Flamethrower Girl as the user profile. Pretty solid. Well done, Ideogram. I definitely don’t think you should sleep on this one; I’ll let you know as it works into my workflow.
Eleven v4 and v4 Turbo
Rounding out the audio side, ElevenLabs has released Eleven v4 along with a Turbo version for real-time voice and agents. Dialogue and sound have been a bane of AI video, so it’s nice to see ElevenLabs, to be honest, kind of catching up with Seed Audio 1.0. With v4 you can put natural-language direction directly into the script, and ElevenLabs says the new model is better at following it, including things like laughter and sound effects.
I returned to a test I think we ran with Seed Audio 1.0: the opening of a high-fantasy audio drama, set at a watchtower. This is another one I’ll be digging into more. I think audio references in Omni video models are going to be pretty big, and it’s taken a while for video models to catch up on audio. Eleven v4 is available now, including on the free tier.
Wrap-up: AI fall
Impressive updates across the board. It’s starting to get a little bit toasty here in AI fall, and the leaves haven’t even started changing color.