LLM knowledge cutoffs: Claude Jun 2026, GPT-6.1 Sol Apr 2026

The Claude 5.5 family lists a June 2026 knowledge cutoff, GPT-6.1 Sol lists April 30, 2026, and Haiku 4.5 lists February 2025. Today is October 3, so even the newest cutoff is months behind. For a video agent that has to know what is trending this week, the model's training data is the wrong source; the agent needs a tool that reads current data.

Sume ships one for TikTok: a trending-videos search that returns ranked public video metadata. It is a research utility, not a generation model, and it is the part of the stack that carries freshness.

What cutoffs do the vendors publish?

Anthropic's models overview separates a reliable knowledge cutoff from a training data cutoff. OpenAI's model page gives one date. The table copies what each page says.

What does the trending-videos tool return?

The Sume doc describes a POST /v1/trending-videos/search call. The platform is TikTok only. A window field takes yesterday, this-week, this-month, last-3-months, last-6-months or all-time, and an optional two-letter region narrows it. Each result carries a public watch URL, author handle, metrics and relevance, and an optional short summary. It does not download or mirror videos.

The doc also notes that the rebuilt browse feed is on for the dev environment and off in production until a flag is set, and that production requires a query and limits limit to 1 to 50. Read the doc for the current contract before building on it.

How does it fit with the model choice?

The cutoff gap is a reason not to prompt for trends from memory in a Format. Put the trend step on a tool, pass its output into the model as input, and let the orchestrating LLM do what it is for: structure, script and shot choice. The model field of a Format run picks only that orchestrator; the call doc says image, video and audio models come from the Format's tools.

That also means the cutoff differences between Claude and GPT-6.1 Sol are a small factor once the trend data arrives as input. Pick the model on cost and context, which the vendor pages list, and keep freshness in the tool.

What should a trend Format do?

A trend Format built this way stays correct as the cutoffs move, because the freshness comes from the tool.

• Search first with a narrow window, such as this-week, and a specific query.
• Hand the returned metadata to the LLM as input; do not ask it what is popular.
• Keep the output tied to the URLs the tool returned, so a reviewer can open each source video.

Is a cutoff ever the deciding factor?

Sometimes. If a script refers to a product, event or tool released after April 2026, a model with a June cutoff is more likely to recognize it than GPT-6.1 Sol, and a model with a February 2025 cutoff such as Haiku 4.5 will not. That is a reason to supply the facts in the input rather than a reason to switch models for every Format.

Anthropic separates the two dates for a reason: the reliable cutoff is where the model's knowledge is dependable, and training data can run later and thinner. For planning, use the earlier date.

None of the vendor pages promise knowledge of a specific trend or video, so treat any claim about a current trend from an unaided model as unverified until a tool or a person has checked it.

One more limit is built into the tool itself: the Sume doc says trending-video search returns metadata and, in the current build, does not download or mirror videos, and summary_mode set to transcript returns metadata plus an unsupported warning. Plan the Format around links and metrics, not around a transcript of each trending clip.