Knowledge cutoff for a trend-research agent: Astra Apr 30 vs Haiku Jun

Short answer

GPT-6 Astra's knowledge cutoff is April 30, 2026, per the OpenAI model page read on 2026-10-08. Claude Haiku 5.5's reliable knowledge cutoff and training data cutoff are both June 2026, per Anthropic's models overview. Both are months before this week's launches, so an agent asked what is trending in short video on October 8 cannot answer from model memory with either. Z.ai's pricing page and the Mistral Large 4 page do not state a cutoff, so no claim is made for those.

Sume lists Haiku 5.5 and Astra in its registry, which gives you a choice, but the cutoff difference is about two months and rarely decides anything for a research task.

The comparison

Cutoffs are vendor statements. Sume does not add its own.

What fills the gap

For anything dated after the cutoff, the agent has to read it from a tool: a web page, a reference video, a platform's rules. Sume's agent has tools and a sandbox, per the Agent Completions docs, and attachments accept up to 30 images. Put the facts that matter, such as a platform's current clip length limit, in the instruction or in input; the docs say Sume writes input to /workspace/inputs/sume-action-input.json and treats it as data, not instructions.

Dated claims are where a stale cutoff hurts: new model names, price changes, and platform rules.

Practical rule

Choose by cost and latency, not cutoff. Haiku 5.5 is Anthropic's fastest listed model and about 100 times cheaper per turn than Astra at the shapes in the related posts. Choose Astra when the planning is hard, and give either one the current facts.

A tiny prompt pattern

One reliable way to deal with cutoffs is to state today's date and the facts that depend on it. The instruction can say: 'Today is 2026-10-08. Platform rules below are current. Do not rely on memory for anything dated after April 2026.' That costs a few dozen tokens and removes a failure mode that no model choice fixes.

For a trend agent, also give it the sources to read, not just a topic. A reference clip or a page URL the agent fetches is better evidence than either model's memory, and the two cutoffs differ by about two months, which is smaller than the lag between a trend appearing and appearing in any training data.

Because the cutoffs are stated by the vendors and the difference is about two months, I would not choose between the two models on this basis alone. The price gap per turn is far larger than any freshness advantage, and the instruction pattern above closes most of the freshness gap anyway. Where freshness really matters, such as a platform that changed its upload limits last month, the answer is to fetch the current page and pass the limits in as data, whichever model you run.

• State the date in the instruction.
• Pass platform limits in input; Sume treats it as data.
• Attach up to 30 images as visual references.