Grok 4.20 Multi-Agent: Parallel Research, Billed Per Agent

Grok 4.20 Multi-Agent looks, on a price list, identical to Grok 4.20: the same $1.25 per million input tokens and $2.50 per million output. That comparison is misleading. One request to the multi-agent variant launches either four or sixteen agents, every one of them consumes tokens, and all of those tokens are billed.

A note on names: xAI, the company behind Grok, was renamed SpaceXAI in July 2026 after SpaceX acquired it. This guide uses the xAI name throughout.

Used for the right job, that is a reasonable trade. Used as a drop-in replacement for a normal chat model, it is an expensive way to get one answer. This guide covers how it works, the restrictions that matter, and how to decide whether it earns its keep.

How the multi-agent variant works

xAI's documentation describes it plainly: multiple agents are launched to discuss and collaborate on your query, and a designated leader agent synthesises the discussion and presents the final answer. Each agent brings its own reasoning and findings. The model page summarises the intended use as "deep research tasks".

From the caller's side it is a single request. You send a prompt and get one response back. By default only the leader's final answer and its tool calls are returned. The sub-agents' state is encrypted and only included when you explicitly ask for encrypted content, which xAI says is there to preserve context across a multi-turn conversation rather than to let you read the sub-agents' working.

The variant shares the Grok 4.20 snapshot, grok-4.20-multi-agent-0309, with a 1M-token context window and text and image input. xAI marks it as beta. For the single-model reasoning and non-reasoning endpoints, see the Grok 4.20 guide.

Four agents or sixteen

There are exactly two configurations. On xAI's own SDK you set agent_count to 4 or 16. Through the OpenAI SDK, the Vercel AI SDK or plain REST, the reasoning effort field selects the count instead: low or medium gives four agents, high or xhigh gives sixteen.

That second mapping is worth writing on a sticky note. On every other Grok model, effort controls how long one model thinks. Here it controls how many agents run. Code that sets high effort everywhere by default will quietly quadruple the fan-out on this model.

xAI's own guidance is four agents for quick, focused queries and sixteen for deep, multi-faceted research, with the explicit warning that more agents mean more tokens and more latency.

What it cannot do

The restrictions are where most integration time goes, so read them before you build:

• No client-side function calling. xAI states that client-side tools and custom tools are not supported on this variant. Your own functions, the core of most agent setups, are off the table.
• Server-side tools only. Built-in web search, X search, code execution and collections search work, as do remote MCP tools. The server runs that tool loop itself.
• No Chat Completions on xAI's API. It works through the Responses API or xAI's SDK only. A gateway that exposes it through a chat-completions endpoint is translating, so check what survives the translation.
• No max_tokens. You cannot cap output length on a single request, which removes one of the usual cost guards.
• Lower rate limits. 9 requests per second and 2.5 million tokens per minute at the time of writing, against 37 and 10 million on the single-model endpoints.

Taken together, this is not a general agent runtime. It is a research box: you give it a question and a set of built-in tools, and it comes back with a synthesised answer.

Reading the bill

xAI says all tokens consumed by the leader and the sub-agents are billed, including input, output and reasoning tokens, and that server-side tool calls from any agent count towards tool usage, which carries its own charge. Its docs warn that one multi-agent request may use significantly more tokens and tool calls than a single-agent one.

xAI does not publish a typical multiplier, and it will vary with the question. Measure it rather than guessing. The response includes usage and server_side_tool_usage fields. Run ten representative questions at four agents, ten at sixteen, and the same ten on single-agent Grok 4.20 reasoning with web search enabled. Divide total cost by the number of answers you would actually use.

Also remember the context-tier rule shared across the family: once a prompt reaches 200K tokens, every token in that request is billed at double the base rate. With several agents each reading a large prompt, that threshold is expensive to cross. Batch requests are supported at a 20% discount, which helps for overnight research jobs.

Built-in fan-out or your own orchestrator

The alternative is to build the pattern yourself: one orchestrator call that splits the question, several worker calls in parallel, and a final synthesis call. Orchestrator-worker patterns covers how to do that well.

Doing it yourself costs engineering time but buys control. You choose the worker model, so the workers can be cheap. You can use your own tools. You can see every intermediate result, cap each worker's output, and retry one failed branch instead of the whole request. The managed variant gives you none of that, in exchange for one API call and no orchestration code.

A reasonable rule: use the managed variant to find out whether parallel research improves your answers at all. If it does, and volume grows, the cost visibility of your own orchestrator usually becomes worth the build. Multi-agent systems: when they help covers the cases where parallelism genuinely improves quality, and agent fan-out limits covers where more agents stop helping.

When to reach for it

• Good fit: open-ended research questions answered from the web or X, competitive analysis, literature sweeps, and questions where checking several angles catches errors a single pass would miss.
• Poor fit: anything that needs your own tools, latency-sensitive interactive work, high-volume short requests, and any workflow where you must cap output per call.
• Start at four agents. Move to sixteen only for questions where you have seen four agents miss something that matters.