Tool support requires the appropriate API integration; a supported tool is not automatically active in every chat.
Supported API features and tools
Coordinated research with 4 or 16 agents and a leader that synthesizes the result
A 1,000,000-token context with text/image input and text output
Provider-built tools and remote MCP tools through supported integrations
Multi-turn research using previous_response_id
Structured outputs and Batch support listed on the model card
Before you choose
Chat Completions is not supported. Use the Responses API or xAI SDK for the documented multi-agent workflow.
Client-side function calling and custom tools are not supported by this variant, despite the generic model card's function-calling badge. The dedicated guide distinguishes supported built-in and remote MCP tools.
The max_tokens parameter is not supported. A separate maximum output ceiling and training cutoff are not verified here.
Only the leader's output is exposed directly. Sub-agent state is encrypted and requires the documented opt-in handling; do not promise readable transcripts from every agent.
Beta interfaces can change. More agents can use substantially more tokens and tool calls without guaranteeing a better answer.
Agent count and effort
lowmediumhighxhigh
Here effort selects agent count, not ordinary reasoning depth: low or medium uses 4 agents; high or xhigh uses 16 agents. The xAI SDK also exposes agent_count with values 4 or 16. The reviewed guide does not establish a default, so choose explicitly when controlling API cost and latency.
Use grok-4.20-multi-agent-0309 to identify the documented variant; grok-4.20-multi-agent is a listed alias. Responses uses reasoning.effort to select the agent setup.
Research can run without built-in tools, in which case agents use their supplied context and model knowledge. Do not describe that as live web research.
Built-in tools must be enabled. Review remote MCP permissions and restrict actions; a supported tool connection is not blanket authorization.
The leader returns the final answer and its tool calls. The xAI SDK's use_encrypted_content option preserves encrypted sub-agent state for supported multi-turn use, not a readable audit trail.
The card lists Batch support, but the dedicated research flow documents Responses and xAI SDK behavior. Verify the precise Batch request and tool combination before depending on it.
Put it to work
Start with a more useful prompt.
Original examples from EZ Ai Assist. Adapt these to your task and the features available in your workspace.
Workflow 01
Compare suppliers with evidence
Set research boundaries and make every comparison auditable.
Compare these three suppliers against the requirements and budget below. Use official product documentation and dated pricing sources if web tools are available; otherwise use only the sources I supply and state that limitation. Compare capability, integration effort, support and unresolved risks in one table. Cite each material claim, flag conflicting evidence and finish with questions to verify before purchase.
Workflow 02
Research architecture trade-offs
Ask for multiple perspectives and a reasoned synthesis.
Evaluate these two architecture options for the workload, team size and operational constraints I provide. Consider reliability, development effort, security boundaries and total operating cost from separate perspectives, then reconcile the trade-offs into one recommendation. Cite supplied evidence or available primary sources. Distinguish assumptions from facts and propose a small experiment to test the highest-risk assumption.
Workflow 03
Reconcile competing product claims
Separate source-backed differences from promotional statements.
Investigate the conflicting product claims below using the supplied sources and, only if available, configured search tools. Build a claim-by-claim evidence table with source date, relevant excerpt and confidence. Explain whether differences reflect version, region, plan or unresolved ambiguity. Return a concise synthesis and a list of claims that still require direct confirmation; do not fill gaps with guesses.
Developer reference
xAI API pricing
These are global xAI API reference prices, not EZ Ai Assist subscription prices.
At 200,000 prompt tokens or more, the higher rates apply to the entire request. These are per-token rates, not a fixed price for a research task.
All billable token usage from the leader and sub-agents counts, including reasoning. Tool invocations or fetched items from every agent also contribute; 16-agent research can cost substantially more than 4-agent research.
Web Search and code execution have call-based fees; X Search bills each fetched post/profile, not each search call. Remote MCP has token usage charges from xAI; a connected service may have its own terms.
The model card lists a 20% Batch token discount. Verify endpoint and tool compatibility for the exact request. The table shows standard global processing, not a promise of every service-tier combination.
Inspect usage and server_side_tool_usage to understand the cost of a completed API task. EZ Ai Assist plan terms and exposed controls are separate.
Common questions
A few things worth knowing.
What makes Multi-Agent different?
It coordinates several agents and a leader that synthesizes their work. This can suit research with competing perspectives or several evidence sources. It is a separate beta variant, not ordinary Grok 4.20 with a longer prompt.
How do I choose 4 or 16 agents?
On Responses, reasoning.effort low or medium selects 4 agents; high or xhigh selects 16. The xAI SDK exposes agent_count as 4 or 16. The reviewed guide does not establish a default. These settings control agent count, not the usual reasoning-depth scale.
Does it always search the web?
No. It can operate without built-in tools using context and model knowledge. Live research requires configured tools such as Web Search or X Search. Ask it to disclose when sources are unavailable and verify cited evidence yourself.
Can I use Chat Completions or custom functions?
The dedicated guide says Chat Completions, client-side/custom function tools and max_tokens are unsupported. Use Responses or the xAI SDK. Built-in tools and remote MCP tools are documented separately; a generic model-card badge does not override these restrictions.
Can I read every agent's intermediate work?
Only the leader's final response and tool calls are exposed directly. Sub-agent state is encrypted and can be preserved through the documented SDK opt-in. That does not provide a readable transcript of every agent's reasoning or work.
How is multi-agent research billed?
Usage from the leader and sub-agents is billed, including reasoning and tool usage. The published token rates match the other 4.20 variants, but a task can consume much more. At 200,000 prompt tokens, higher rates apply to the request; choose the agent count deliberately.
Are the beta features available in EZ Ai Assist?
This guide does not establish that. xAI documents a beta API with a 1M context and potentially changing behavior; separate output and cutoff limits are not verified here. Check EZ Ai Assist for model access, tools, agent controls and plan terms.
Check the source
Official documentation
Specifications and API prices checked on . Example prompts and workflow advice are editorial guidance from EZ Ai Assist.