Can Fable 5.1 and GPT-6 Astra make video?

The same sentence sits under the videos going round social media: "AI made this in one click". The models named are usually the same two: OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1. So ask it plainly: can these two make video?

Short answer: no. Not because it is switched off, but because that output modality does not exist for them.

What the documentation says

GPT-6 Astra. OpenAI's model page marks video as an unsupported modality. Audio too. Images can come back, but the model is not the renderer: a separate image-generation tool is called through the Responses API.

Claude Fable 5.1. Anthropic's docs put it in one sentence: all current models take text and image input and produce text output. Never mind video, there is no image output either.

OpenAI's video model is being retired

Here is the part that often gets skipped. The basis for "OpenAI does video" was Sora, and Sora is closing. The timeline:

• 24 March 2026 · the shutdown was announced.
• 26 April 2026 · sora.com and the app went dark.
• 24 September 2026 · the API closes. Two weeks from now.

Content tied to the account is to be deleted after the cutoff. OpenAI has given no timeline for video generation returning. The reason reported is economics rather than capability: running costs never matched revenue, early interest fell away, and copyright and likeness questions stayed unresolved.

So if the question is "does OpenAI have a video model", the answer changes at the end of this month.

What these models actually do, and how

This is the real question: where do they enter the chain, by hand or through an API? There are two routes, and in neither does the model render the picture.

1. By hand, writing prompts. You ask the model to write a video prompt for a shot, then copy that text into the video tool's own interface. There is no connection between the two products; you are the carrier. This is exactly where Astra's "I only write prompts" sits, and it is what most people do.

2. Through an API, as a tool call. You define a tool for the model: a "generate video" function with the video model's API behind it. Mid-conversation the model decides to call it, writes the prompt itself, sends it, and hands you the file that comes back. On OpenAI's side this is the Responses API tool mechanism; on Anthropic's it is tool use, or MCP.

The "one button" in the demos is this second route. But note what is being called: somebody else's API. The model still writes, decides and triggers; it is not the thing laying down frames. The only difference between the routes is who carries the text.

Its other jobs in the chain

• Script and shot plan. Breaking the subject into beats, setting shot lengths and order.
• Shot prompts. Camera angle, light, movement, duration, translated into the language the video model understands. Usually the link that decides quality more than any other.
• The chain's code. The script wiring the video model, the voice step and the editing together. The "one button" is usually that script.
• Reviewing the frame. Because they accept image input, you can show them a rendered frame and have them say "the hand has six fingers, fix the prompt like this".
• Text work. Title, description, subtitles, translation.

In short these models write, decide and trigger. They do not render.

Who renders the picture

Separate models lay down the frames: Seedance (ByteDance), Runway, Kling, Veo, Luma. Where a 3D scene is built, a classic package like Blender joins in. Voice is another model again. And there are open-weight video models such as Wan that run on your own hardware.

Most of the people making these videos already say so: the tool is named in the description or the end credits. Nothing is being hidden.

What the comparisons on X measure

This bears directly on the comparison videos. Frames placed side by side under "Astra made this, Fable 5.1 made that" are not the output of two language models; since neither can produce video, a third model made the picture.

There is no contest between two models here, but between two chains. The video link in each can even be the same product, in which case the difference comes from who wrote the better prompt. An honest comparison states which model rendered the video and which settings were used.

What this actually is

Looking for deception would be wrong; the tools are usually named. The problem is what sticks: the headline carries two language models' names, while the model that made the picture sits down in the description.

Sora's shutdown teaches something else. A video model you rent can be switched off, and not for lack of capability but because the numbers did not work. A team that built its chain on a single closed product has to rebuild that link when the product closes. Open-weight models running on your own hardware are therefore not only a cost question but a continuity one.

The right question: how many links does the chain have, which model owns each one, and what do you do if that model closes tomorrow.

Sources: Anthropic model docs, OpenAI GPT-6 Astra, OpenAI on the Sora discontinuation.