What is Grok 4.20 Multi-Agent?
Grok 4.20 Multi-Agent is xAI's team-of-agents model in the Grok 4.20 line: a leader agent hands the question to four or sixteen sub-agents that research and analyze in parallel, then writes the final answer from their work. It went live on the API on March 10, 2026 alongside Grok 4.20 and is still labeled a beta. It accepts text and image input, outputs text, and runs on Unifically as xai/grok-4.20-multi-agent-0309 with a 1M-token context window. One thing sets it apart on the wire: xAI serves it only on the Responses API, so you call /v1/responses, not chat completions.
Key features of Grok 4.20 Multi-Agent
Four or sixteen agents per request
Set `reasoning.effort` to `low` or `medium` for a 4-agent team suited to focused questions, or `high` and `xhigh` for a 16-agent team built for deep, many-sided research. The leader synthesizes the result; you get one answer, not sixteen.
Parallel research in one call
Each sub-agent can pursue its own thread of the problem at the same time, which is the point: broad questions get covered from several angles inside a single request instead of a hand-rolled agent loop.
Human voters rate it with Grok 4.5
On the Arena text leaderboard the 0309 multi-agent beta sits at #37 with an Elo of 1470 across 60,846 votes as of September 2, 2026, level with Grok 4.5 and ahead of Grok 4.6 high.
A million tokens of shared context
The whole team works from the same 1M-token input, with image understanding for charts and screenshots.
Best for
Deep research questions
Sixteen agents cover a many-sided topic in one request.
Competitive and market analysis
Parallel threads compare options and synthesize a view.
Due-diligence summaries
Long inputs, many angles, one written answer with the leader's judgment.
Second-opinion answers
Agents debate a hard question before the leader commits.
Long-document synthesis
1M tokens of context shared across the team.
Chart and screenshot analysis
Image input feeds the same multi-agent pass.
Use cases
Use Grok 4.20 Multi-Agent for the questions you would otherwise assign to a small team: market landscapes, technical comparisons, policy reviews, and any research brief where breadth matters and a single model tends to miss angles. Feed it long source material, up to a million tokens, and let the sixteen-agent setup work the problem; drop to four agents for quicker, narrower asks. It also works as a slow, careful second opinion on decisions a fast model already drafted. Because it runs only on the Responses API, wire it through /v1/responses and keep chat completions for the rest of the Grok line.
Limitations
It is a research model, not a chat model. Sixteen agents take longer and consume far more tokens than one, and every agent's tokens are billed, so cost per request is a multiple of Grok 4.20 Reasoning.
You control less. Client-defined tools are not supported, max_tokens is rejected, and the sub-agents' intermediate reasoning is not returned, only the leader's final answer.
It is chat-completions-incompatible by design: /v1/chat/completions returns an upstream error for this model ID. The playground on this page talks to /v1/responses for that reason.
Arena voters place it at #37, level with Grok 4.5 rather than above it, so the team does not buy frontier intelligence; it buys coverage. And it is a beta, with breaking changes possible.
Grok 4.20 Multi-Agent vs Grok 4.20 Reasoning
Reasoning is one model thinking once; Multi-Agent is the same base run as a team. Arena voters score them within two Elo points (1472 against 1470), so for a single well-posed question Reasoning gives you a similar answer faster and cheaper, with tools and streaming available. Pick Multi-Agent when the question is wide, the sources are long, and you would otherwise build a fan-out agent loop yourself.
When to use Grok 4.20 Multi-Agent
Use it for research-shaped requests where breadth and synthesis matter more than latency or price, and where you can call the Responses API. Skip it for interactive chat, tool-driven workflows, and anything cost-sensitive; those belong on Grok 4.20 Reasoning or Non-Reasoning.
API examples
Call Grok 4.20 Multi-Agent by POSTing to /v1/responses, the OpenAI-compatible Responses endpoint. This model is not served on /v1/chat/completions. Full parameter docs live at docs.unifically.com/models/llm/xai/grok-4.20-multi-agent-0309.
curl -X POST https://api.unifically.com/v1/responses \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "xai/grok-4.20-multi-agent-0309",
"reasoning": { "effort": "high" },
"input": "Compare the three leading approaches to on-device speech recognition and recommend one for a battery-constrained wearable."
}'
The response comes back synchronously with the leader agent's answer in output_text. Use "effort": "low" for the 4-agent setup and "high" for 16 agents.
FAQs
People also ask
xai/grok-4.20-multi-agent-0309, called through the OpenAI-compatible POST /v1/responses endpoint with your Unifically API key. xAI does not serve this model on chat completions, so /v1/chat/completions returns an error for it.
A leader agent farms the question out to a team of sub-agents that research and analyze in parallel, then synthesizes their work into one answer. You see the leader's output; the sub-agents' intermediate reasoning stays encrypted upstream.
Four or sixteen. Set reasoning.effort to low or medium for the 4-agent setup, or high or xhigh for the 16-agent deep-research setup.
Text and image input, text output, with a 1M-token context window. Function calling and structured outputs are supported upstream, but client-defined tools are not, and max_tokens is not accepted.
Yes, per request. Every token consumed by the leader and each sub-agent is billed, including reasoning tokens, so a 16-agent call uses far more tokens than a single Grok 4.20 call on the same prompt.
The 0309 multi-agent beta sits at
Yes. xAI labels it a beta and reserves the right to change how it works, so pin your prompts to the 0309 ID and re-test after upstream updates.
