> ## Documentation Index
> Fetch the complete documentation index at: https://docs.apiyi.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Grok 4.20 Beta Series: 4-Agent Architecture, 2M Context
> xAI releases Grok 4.20 Beta with 4 models featuring a multi-agent collaboration architecture, 2M token context window, and unified pricing at $2/$6 per million tokens. Now available on APIYI.
## Key Highlights
* **4-Agent Architecture**: Grok 4.20's core innovation — 4 specialized agents think in parallel and debate before producing answers, reducing hallucinations by \~65%
* **2M Context Window**: 8x increase from Grok 4's 256K to 2 million tokens
* **4 Models for Every Use Case**: Reasoning, multi-agent, base, and non-reasoning variants for everything from instant responses to deep research
* **Unified Low Pricing**: \$2 input / \$6 output per million tokens, a major price drop from Grok 4 (\$3/\$15)
* **Multimodal Input**: Supports text + image (JPG, PNG) input
## Background
On February 17, 2026, xAI launched Grok 4.20 Beta on `grok.com` for consumers. The API variants were released on March 9-10, 2026. Grok 4.20 represents a major architectural shift from a single model to a multi-agent collaboration system.
The core innovation is a 4-agent architecture: Grok (captain/coordinator), Harper (research & fact-checking), Benjamin (logic/math/coding specialist), and Lucas (creative synthesis & devil's advocate). These agents work in parallel, debate each other, and produce more accurate, comprehensive answers.
All 4 models are now available on APIYI with pricing matching xAI's official rates.
## Detailed Analysis
### Model Comparison
| Model | Purpose | Best For |
| - | - | - |
| `grok-4.20-beta` | General purpose | Daily conversation, content generation, general tasks |
| `grok-4.20-beta-0309-reasoning` | Enhanced reasoning | Complex logic, multi-step math, scientific reasoning, coding |
| `grok-4.20-beta-0309-non-reasoning` | Ultra-fast | Low-latency scenarios, simple Q\&A, classification |
| `grok-4.20-multi-agent-beta-0309` | Multi-agent | Deep research, complex multi-step workflows |
### Core Features
Grok (coordinator), Harper (research), Benjamin (logic), Lucas (creative) — parallel thinking with debate
2 million tokens, 8x larger than Grok 4's 256K
Multi-agent collaboration reduces hallucination rate from \~12% to \~4.2%, a \~65% improvement
\$2 input / \$6 output per million tokens, \~60% cheaper than Grok 4
### Performance Data
Data sourced from Artificial Analysis and other third-party benchmarks. xAI has not yet published official comprehensive benchmark results.
| Metric | Reasoning | Non-Reasoning | Grok 4 (Reference) |
| - | - | - | - |
| **AA Intelligence Index** | 48 | 30 | 42 |
| **Output Speed** | \~231 t/s | \~232.5 t/s | — |
| **Context Window** | 2M | 2M | 256K |
| **Input Price** | \$2/M | \$2/M | \$3/M |
| **Output Price** | \$6/M | \$6/M | \$15/M |
### Multi-Agent Model Details
`grok-4.20-multi-agent-beta-0309` includes built-in tool capabilities:
* **web\_search**: Web search
* **x\_search**: Real-time X (Twitter) data search
* **code\_execution**: Code execution
* **collections\_search**: Knowledge base search
API calls return only the lead agent's final response. Sub-agent intermediate reasoning is not exposed by default.
## Getting Started
### Code Example
```python theme={null}
from openai import OpenAI
client = OpenAI(
api_key="your-api-key",
base_url="https://api.apiyi.com/v1"
)
# Reasoning variant - for complex logic tasks
response = client.chat.completions.create(
model="grok-4.20-beta-0309-reasoning",
messages=[
{"role": "user", "content": "Analyze the time complexity of this algorithm and suggest optimizations..."}
]
)
print(response.choices[0].message.content)
```
```python theme={null}
# Non-reasoning variant - for low-latency scenarios
response = client.chat.completions.create(
model="grok-4.20-beta-0309-non-reasoning",
messages=[
{"role": "user", "content": "Translate the following text to Chinese: ..."}
]
)
print(response.choices[0].message.content)
```
### Recommended Use Cases
Math competitions, scientific analysis, complex coding, multi-step logic
Quick Q\&A, text classification, translation, data extraction
Daily conversation, content creation, general assistant
Deep research reports, complex investigations, multi-perspective analysis
## Pricing & Availability
### Pricing
| Model | Input Price | Output Price | Billing |
| - | - | - | - |
| `grok-4.20-beta` | \$2 / M tokens | \$6 / M tokens | Pay-per-use |
| `grok-4.20-beta-0309-reasoning` | \$2 / M tokens | \$6 / M tokens | Pay-per-use |
| `grok-4.20-beta-0309-non-reasoning` | \$2 / M tokens | \$6 / M tokens | Pay-per-use |
| `grok-4.20-multi-agent-beta-0309` | \$2 / M tokens | \$6 / M tokens | Pay-per-use |
### Deposit Bonuses
Top-up promotions apply. See [promotion details](/en/faq/recharge-promotions).
## Summary & Recommendations
Grok 4.20 Beta's standout feature is its 4-agent collaboration architecture and 2M ultra-long context. While its AA Intelligence Index (48) trails Gemini 3.1 Pro (57) and GPT-5.4 (57), the unique multi-agent mechanism excels at reducing hallucinations and improving accuracy on complex tasks. The unified \$2/\$6 pricing is highly competitive.
Grok 4.20 is still in Beta. xAI states they are shipping bug fixes and improvements daily. Implement proper error handling for production use.
Sources: xAI official documentation at `docs.x.ai`, Artificial Analysis benchmark data, xAI release announcements. Data retrieved: March 2026.
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.