At this stage we have consulted no primary source for this announcement: details on pricing, supported languages and terms of use therefore remain to be verified.

What it is for

This kind of tool meets a very concrete need for short-form content creators: producing explainer videos, product presentations or series with a recurring character, without filming or frame-by-frame animation. The simplicity of the input (one image, one text) is the main selling point.

Questions to ask

A tool that can make any image talk raises image-rights questions. Before using a real person’s face, you need their consent. And for commercial distribution, the service’s license terms must be checked.

To watch: CapCut in ChatGPT

The same source announces the upcoming arrival of a CapCut plugin in ChatGPT. With no date and no official confirmation consulted, we list it among the information still to be confirmed.

Cursor: the agent can use a computer

On October 23, Cursor published documentation for Computer Use in its self-hosted cloud agents. The idea: the agent is no longer limited to files and the terminal; it can also drive a graphical environment, for example to check how an interface renders. The fact that the feature is documented in the “self-hosted” section is notable: the company keeps control of the machine on which the agent acts.

Claude Code: more than ten sandbox flaws fixed

On the Claude Code side, the changelog for versions 2.1.309 and 2.1.310 reports fixes for more than ten bypasses of the sandbox and permission system. In other words, there were more than ten ways to get outside the perimeter the agent was allowed to work in.

Why this is one story

The more capabilities an agent has (running commands, clicking, browsing), the more the quality of its isolation matters. A bypassed sandbox on an agent that drives an entire computer has far heavier consequences than on a simple autocomplete assistant.

Two protocols, two results

The gap does not necessarily mean one of the two figures is wrong. According to the ARC Prize, OpenAI’s score was obtained with an adapted protocol, different from the one applied to every model it evaluates. Yet it is the common protocol that makes it possible to compare systems with one another. The reference figure is therefore 62.7%.

Cost, the other key figure

The ARC Prize also states the cost of its evaluation: $26,098. That amount is a reminder that a reasoning score is paid for in compute, and that two models with the same score can have very different costs. A result without its cost tells only part of the story.

A price that changes the equation

The model is billed at $2 per million input tokens and $10 per million output tokens. For a model ranked among the top four, that is aggressive positioning: high-volume applications, such as agents that run long sessions back to back or document processing, are the ones where the per-token price weighs most heavily on the bill.

Weights promised, not yet released

The developer says the model weights will be released by November 4. If that promise is kept, K3.1 Pro would become one of the highest-ranked open models, downloadable and self-hostable. For now it is only a commitment: we will check when the deadline arrives.

What we know, what we don’t

General availability is the solid fact in this story. It means developers and businesses can now build on Argon without a waiting list, which will also allow the first large-scale independent evaluations.

On the other hand, the existence of a “Carbon” variant rests solely on a Business Insider article, according to which Google employees are testing new Gemini 4 models. Google has not confirmed it. We are therefore treating it as a rumor: neither its specifications, nor its timeline, nor even its release are certain.

Why the distinction is useful

Code names that circulate before a launch sometimes become products and sometimes go nowhere. For a team choosing a model today, the only actionable information is that Argon is available. Waiting for a hypothetical next version only makes sense once it has been officially announced.

The figure does not fully survive testing

The Reel Stack ran the numbers on its own. The result: savings of 26%, not 62%. The gap is large, but it does not make T1 a bad deal: cutting costs by a quarter is still significant for anyone who produces a lot of video.

This kind of gap is common with credit-based pricing. The real cost depends on how credits convert into euros under each subscription plan, on clip length, and on how many failed generations have to be rerun. A displayed “per second” comparison says nothing about any of that.

Seedance moves upmarket

Meanwhile, ByteDance launched Seedance 2.5 4K on October 23, aimed at film production, according to TechNode. This version is restricted to certain companies.

Why keyframes matter

Many video generators work from a text prompt and, at best, a start frame and an end frame. The result is often spectacular but hard to direct. With ten keyframes, users can pin down intermediate steps: a character entering, a change of scenery, the final framing. It gets closer to a storyboard that the model then animates.

The 30-second length points the same way. A shot of a few seconds is enough for an effect; thirty seconds allow for a short scene, a product demo or a complete narrative transition.

The developer’s announcement stresses three themes: visual realism, creative control and storytelling. The first two are a matter of judgment and have not yet been the subject of published independent measurements.

Nano Banana 2.1, released on October 6 by Google DeepMind, accepts up to 14 reference images at a cost of about $0.034 per 1K image. Krea 3, launched on October 15, offers 2K editing, up to 6 references, at $0.03 per 1K image.

What the numbers say

The price gap is small: about 12% in Krea 3’s favor at 1K. Over a volume of 10,000 images, that comes to around forty dollars. The real deciding factor lies elsewhere:

• Many references to combine (a product line, several characters): Nano Banana 2.1 accepts more than twice as many.
• Resolution needed for editing: Krea 3 highlights 2K editing.

Two promises stand out in the announcement. First, native 4K generation, with no upscaling step. Second, an INT8-quantized version that would fit in 8 GB of video memory, the class of consumer graphics cards most people actually own. Both are specifications announced by the developer: the real quality of the 4K output and the speed on 8 GB have yet to be measured.

A license designed for small businesses

The other novelty is commercial: the license is free for commercial use as long as the company makes less than $5 million in revenue. Freelancers, agencies and startups can therefore build it in at no cost, while larger players will need to sign an agreement.

The model comes in two variants: Base, the pre-trained model, and SFT, refined through supervised fine-tuning. The weights are published on Hugging Face and the code on GitHub, so anyone can audit it and fine-tune it themselves.

A ranking still to be confirmed

According to the team’s report, MOSS-Image ranks 2nd on RealLayoutArena, a benchmark focused on layout. That is promising for uses where composition matters (posters, visuals with text), but it is a self-reported result: we are waiting for an independent measurement.

Already an editing add-on, with caution

A secondary source reports a community add-on, Moss-Image-Edit, which lets you edit an image with a single sentence in ComfyUI and is said to run on an 8 GB RTX 4060 in 8-bit. The problem: its weights have no clear license. For professional use, it is better to stick to the official model for now.