Limits and costs
Limits keep research capacity fair and stop runaway agent loops. They apply per Instant Expert account, so every assistant you connect (the ChatGPT plugin included) draws from the same allowance. The numbers on this page are read from the code that enforces them.
Account limits
| Limit | Value | Applies |
|---|---|---|
| MCP requests | 120 per minute | Per account, shared by every connected assistant |
| Searches, imports and drafts started | 50 per rolling 24 hours | Per account, shared by every connected assistant |
| Per-action budgets | search: 5 per minute, 20 per hour, 50 per day; import: 5 per minute, 10 per hour, 30 per day; draft: 30 per minute, 200 per hour, 500 per day; plan: 5 per minute, 30 per hour, 100 per day | Per account, shared by the website and every connected assistant |
| Job stops if it makes no progress for | 15 minutes | Per job |
| Longest a job can run | 45 minutes | Per job |
| MCP request body | 128 KB | Per HTTP request |
| Website searches without an account | 1 per day | Per network (IP address) |
| Anonymous sandboxes per day | 500 per day | Per network (IP address) |
| Anonymous sandboxes per ip per hour | 5 per hour | Per network (IP address) |
| Anonymous search ip limits | 2 per minute, 6 per hour, 10 per day | Per network (IP address) |
When you go over the per-minute limit, the server answers with HTTP 429 and a Retry-After header. When you go over the daily limit, tools that start work return a daily_limit error saying how many seconds to wait. Searches, imports, drafts and outreach plans (plan_outreach) also each have per-minute, per-hour and per-day budgets, shared with the website; going over one returns a rate_limited error with the number of seconds to wait. Reading tools (get_job, get_search, list_requests and the like) don't count toward the daily limit, but they do count toward the per-minute one, so poll at poll_after_seconds rather than in a tight loop.
On the Free plan, an account can also start 100 searches and create 3 new Advisor plays each calendar month (UTC), on the website and every assistant together. Past that, search_people and create_play return a plan_limit error that says when the month resets and links the plans. Plays you already have keep running their batches, and those batches don't count.
Per-call limits
These come straight from the tool schemas (see the tool reference).
| Limit | Value |
|---|---|
Search request length (search_people) | 2,000 characters |
Contacts per import (import_people) | 100 people |
Recipients per draft (queue_requests) | 100 people |
Request message length (queue_requests) | 500 characters |
Minimum a person receives (offer_cents) | $5 |
Most a person receives (offer_cents) | $250,000 |
Largest Advisor budget (total_budget_cents) | $250,000 |
People per get_search page | 100 |
Requests per list_requests page | 50 |
People cards per show_people call (ChatGPT) | 8 |
idempotency_key length | 128 characters |
Costs
Searching, importing and preparing drafts don't cost money; they only count toward the limits above. After an account's first 10 searches, starting another needs a card saved under Settings; until then a search returns a card_required error. Saving the card charges nothing. You pay for requests people accept:
- Calls are charged when the person books a time.
- Written or voice answers are charged when the reply is completed.
- Requests nobody accepts cost nothing.
On Managed, you pay one price per call or answer, our fee included, and we choose what each person is offered. On every plan you set one price per booked call or completed answer, Instant Expert's fee included, and Instant Expert decides what each person is offered from it; checkout shows that one price. Assistants set it as price_cents. The one exception is an account holding a promo code that waives the fee, added under Settings: there you name what each person receives and the fee goes on top (a $100 offer costs you $125, the fee being 20% of what you pay) before the code waives it. It applies to requests sent while the code has uses left and hasn't expired (some codes last a set time after you add them, shown next to the code in Settings), a booked call or paid reply uses one, and a refunded one gives it back. What each person is offered runs from $5 to $250,000, and an Advisor plan's budget can be up to $250,000; the website applies the same limits. Each person also has a minimum of their own, about a quarter of their suggested amount and never more than $32, which queueing and sending enforce. When you send, one authorization hold is placed for the priciest person's price, fee included, and each accepted request is then charged separately, so the most a batch costs is everyone accepting. An Advisor plan's batches also spend from the plan's one budget, first come, first served, at what you pay. Cancellations, no-shows and refunds follow the Terms of Service; How it works explains them in plain terms.
For assistants and LLM tools: this page as Markdown.