Search the web

A model on its own knows nothing after its training cut-off and cannot open a page. Turn on Search the web and, for that turn, it can: it searches, reads what it finds, and answers with numbered sources under the reply. Search and read is all it can do; it cannot act on anything it finds.

Turning it on

The globe sits beside the paperclip in the composer, in Ask and Code. Tap it and it stays on for that chat until you tap it again; each chat remembers its own setting. The estimate beside Send grows by the most the searches can add, and on a phone a line under the tool row says the same: Search the web is on, up to 15 cr search.

On a model that cannot search, the globe stays in place but off, and hovering or focusing it says why: <model> cannot search the web. Pick a model that can. Image and video models never offer it.

What happens in a turn

1. The model decidesIt is given two tools, web_search and fetch_page, and told today's date. It searches when the question needs current, specific or verifiable information and answers directly when it does not, so a turn with the globe on may run no search and cost nothing extra.
2. Searching the web for ...Each search returns the most relevant passages from up to 5 pages, with their titles and addresses, trimmed to about 3,000 tokens. The model can limit a search to the past day, week, month or year. The line above the streaming answer shows the query as it runs.
3. Reading ...A page you named, or one the model needs in full, is read by askr's server as plain text: the main content, trimmed to about 2,000 tokens. The line shows the site. A page whose robots.txt forbids automated reading, a private or internal address, and anything that is not HTML or plain text are refused.
4. The answerEverything a tool returns reaches the model as data between markers, with a standing instruction that nothing inside them is an instruction, so a page that says "ignore previous instructions" is quoted, not obeyed. The model cites what it uses as [1], [2] and so on.

Searches

At most 3 a turn

Pages read

At most 2 a turn

Model calls

At most 4 a turn. The last is made without tools, so the model has to answer with what it has.

Pace

6 searching turns a minute per account, workspace and API together

A search that fails

The model is told, and answers from what it knows, saying so if it matters. It still counts as a search.

What it costs

• 5 credits a search. Charged for every search that ran, whatever it found.
• The text the model reads, as input tokens. Search results and pages go to the model like anything else in the conversation, at the model's own input rate, so a search on an expensive model costs more than the same search on a cheap one. Reading a page costs nothing beyond its tokens.

Before you send, the estimate says the most the searches can add: ~12 cr + up to 15 cr search is three searches at 5 credits. The hold on your balance is those 15 credits plus the most text the searches, the pages and the extra model calls could add at the model's rate, on top of the usual hold for the reply. When the turn finishes the hold settles to what actually ran, and the receipt says so: 2 SEARCHES, 10 CR, with the pages read counted beside it when any were. A turn that searched but produced no answer is not charged, searches included. Holds and settlement.

Sources and the [n] marks

Under a reply that searched is a Sources line: the pages the turn read, numbered in the order it met them, each shown as its domain and opening the page in a new tab, with the title on hover. A [1] in the answer opens the same page as source 1. A page found by two searches has one number. History and Search mark a chat that searched with a small globe.

Which models

Any chat model the catalog marks as taking tools, which is most of them. The public JSON at /api/models says supportsTools per model and carries the search offer (the price and the caps); the API lists the same flag on GET /v1/models. If a provider turns out to reject the tools after all, the turn runs again without them, so you still get an answer, without sources.

Privacy

When Search the web is on, the search terms the model chooses are sent to Brave Search; page fetches are made by askr's server, never from your browser. Your account travels with neither. askr's fetch introduces itself as askr/1.0 (+https://heyaskr.ai) and honours robots.txt. Privacy.

If it will not run

Each refusal ends with a code from the list: W206 when this server has no search, W207 when the model cannot use tools, W109 when the account has started more than six searching turns in a minute, W110 when something other than the workspace asked to search without streaming. None of them charge anything.

The same search runs on the API as web_search: true. Chat completions.