November 12, 2026 · 5 min

Fanout 1.1: the searches an AI assistant runs, with real products and this year’s dates

Fanout 1.1 is out today, and it is now the model behind every Fanout request. If you already call fanout-1, you are on 1.1 from today and there is nothing to change. To name the version, call fanout-1.1. The price has not moved.

Fanout takes a question someone would ask an AI assistant and returns the web searches that assistant is likely to run before it answers. Those searches decide which pages get cited in the answer, so they are the brief for anyone who wants to be one of them.

What is new

• Real products, not plausible ones. Searches name the products and companies that actually exist in a category. Half as many of its searches are rejected by a validity filter as from models 15 to 34 times its size.
• This year’s searches. The current date goes into every request for you, so searches carry this year. Fanout 1 wrote an old year in one search in seven.
• A category hint. Send category, on its own or with the brands you know, and the searches stay in your space. For a brand it has never seen and no hint, it searches for the brand itself rather than guessing what it sells.
• Clean output. Names are spelled the way you wrote them, with no stray quotes, no half-finished “X vs” and no retired model names.
• On GPUs for the API. Paid requests now run on GPUs, at about 0.2 seconds a request.

How it compares

We asked every model for the searches behind 152 buyer questions from 19 product categories that were not in our training data, and compared them with the searches a real assistant, Claude Haiku 5.5 with web search, actually ran for the same questions. Coverage measures how much of what the real assistant searched for a model found.

Fanout 1.1 matches Gemma 4 12B and Qwen3.8-27B: the three sit within each other’s 95% intervals, at a fraction of their size and cost. It covers far more than Fanout 1 and than the open query-expansion models, which are built to broaden a search box query rather than to predict what an assistant looks up. The dashed line is what a second run of the real assistant scores against the first: engines never search the same way twice, so nothing reaches 1.0.

Coverage is half of it. The other half is whether the searches are real:

A search that names a product nothing in the question supports, or wanders off the question, is a search no assistant would run, and a brief built on it sends you after the wrong pages. Fanout 1.1 produces half as many as the larger models.

In production

We also ran 102 real questions from production through Fanout 1 and Fanout 1.1 and scored them with the production validity filter, which knows each brand’s real competitors.

Read by hand, 47 of 50 Fanout 1.1 sets were usable as they came, and none carried an old year, against 64 searches with one from Fanout 1.

Three examples

Questions from categories the model never trained on, with what the real assistant searched. These are the actual outputs, not edited.

Fanout 1 invents “InvoicesAI” and “Gadget World” and searches for 2024. Fanout 1.1 stays on the question and on this year.

Speed

The paid API runs on GPUs with inference tuned for this model; the free playground runs the same model on CPU. One GPU sustains about 11,000 tokens a second across concurrent requests.

All the numbers

On research questions from Microsoft’s public Researchy Questions set, which asks for searches behind open research questions rather than buying decisions, Fanout 1.1 also improves on Fanout 1 (0.358 against 0.343 coverage of the reference searches). General models built for broad knowledge do better there (Gemma 4 12B 0.429, Qwen3.8-27B 0.400). Fanout is built for how people shop.

Using it

curl https://meragpt.com/v1/fanout \
  -H "Authorization: Bearer $MERAGPT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "fanout-1",
    "question": "What are the top alternatives to Pinecone?",
    "category": "vector database",
    "brands": ["Pinecone", "Weaviate", "Qdrant"]
  }'

The response has queries, up to six searches. category and brands are optional. Read the model page and the docs, try it in the playground, or sign up for $1 of free credit, no card needed.

How we measured

New-category benchmark: 152 questions, eight question types (best-of lists, alternatives, leading companies, recommendations, comparisons, popular tools and two kinds of opinion) across 19 categories none of the models were trained on by us. The reference is the searches Claude Haiku 5.5 ran with web search, one run per question; the ceiling is a second run on 120 of them. Every model ran three times at temperature 0.7, and every prompted model was told the same date as Fanout 1.1. Coverage is, for each real search, the best word overlap (F1) with any predicted search, averaged. Intervals are 95% bootstrap intervals over questions. The validity filter is the one our AI-visibility product uses, with only the question’s own names as known products, which makes it stricter here than in production. Production set: 102 real questions, scored by the same filter with full context and read by hand.