Billing note Sub-agent tokens are billed on top of your own prompt and answer, so every turn has a floor of several thousand tokens regardless of input length (measured: a one-word prompt settled 4,266 tokens ≈ 9 credits). Use it for research-grade questions, not quick chat. Chat Completions is rejected upstream (400) - call POST /llm/grok/v1/responses.

xAISlowUltra

Grok 4.20 Multi-Agent

Grok 4.20 Multi-Agent beta (0309 snapshot) - deep-research style multi-agent orchestration with a 1M-token context window. Served through the Responses API only.

2,323/4,645credits

input / output · per 1M tokens

Multi-agent deep research (beta) - sub-agent tokens are billed

1M context window

Responses API only (/v1/responses) - Chat Completions is rejected with 400

Prompt caching with discounted cache reads

Long-context tier: prompts reaching 200K tokens bill the whole request at 2x rates

Run it right now

Test this model instantly in the Console Playground — no code required

Use with AI Assistant

Copy usage instructions for Claude, ChatGPT, or other AI

llms.txtllms.txt ↗

Model Specifications

Context Window

tokens

Max Output

33K

tokens

Training Cutoff

2025-11

Compatible SDK

OpenAI

Capabilities

Vision

Function Calling

Streaming

JSON Mode

System Prompt

Token Pricing (per 1M tokens)

Token Type	Credits	USD Equivalent
Input Tokens	2,323	$1.55
Output Tokens	4,645	$3.10
Cached Tokens	371.6	$0.25

* 1,500 credits ≈ $1 (actual charges may vary based on usage)

Quick Start

curl -X POST "https://api.core.today/llm/grok/v1/responses" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer cdt_your_api_key" \
  -d '{
  "model": "grok-4.20-multi-agent-0309",
  "instructions": "You are a research analyst. Cite the evidence you rely on.",
  "input": "Compare the three leading approaches to long-context retrieval in 2026 and recommend one for a legal-document pipeline.",
  "max_output_tokens": 4096
}'

Parameters

Parameter	Type	Required	Default	Description
`model`	string	Yes	-	Model ID, e.g. "grok-4.20-multi-agent-0309".
`input`	string \| array	Yes	-	Responses API input - a plain string or an array of role/content messages (replaces `messages`).
`instructions`	string	No	-	System-level instructions (replaces the `system` message).
`max_output_tokens`	integer	No	4096	Cap on the visible answer. Sub-agent and reasoning tokens are billed separately and are not limited by this value.
`stream`	boolean	No	false	Stream the response as server-sent events.

Examples

Deep research (Responses API)

Responses API request - the only endpoint this model accepts. A research-grade question amortizes the per-turn token floor.

curl -X POST "https://api.core.today/llm/grok/v1/responses" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer cdt_your_api_key" \
  -d '{
  "model": "grok-4.20-multi-agent-0309",
  "instructions": "You are a research analyst. Cite the evidence you rely on.",
  "input": "Compare the three leading approaches to long-context retrieval in 2026 and recommend one for a legal-document pipeline.",
  "max_output_tokens": 4096
}'

Tips & Best Practices

1Call POST /llm/grok/v1/responses - Chat Completions returns 400 for this model. With the OpenAI SDK use client.responses.create(...) with base_url https://api.core.today/llm/grok/v1

2Budget for a per-turn floor of roughly 4K tokens (≈9 credits): the orchestrator's cached system context and sub-agent reasoning are billed even when your prompt is one line. Pick a single-agent Grok for short Q&A

3max_output_tokens caps only the visible answer - it does not cap sub-agent or reasoning tokens

4Keep prompts under 200K tokens to stay on the standard rate (2x above that)