In February this year, xAI launched Grok 4.20 — not Grok 5, but something structurally different: four role-playing agents running inside the same model, questioning and debating each other before synthesizing a final answer.
The idea sounds abstract, but it rests on a concrete engineering design.
Four roles, one brain
First, the structure: Grok 4.20 is not four separate models but a single MoE base with roughly 3 trillion parameters, activating about 500 billion parameters per inference. The four agents run on this base through lightweight LoRA-style adapters — essentially four "personalities" of the same model.
The four roles are:
- Grok (Captain): task decomposition, overall strategy, final synthesis output
- Harper (Researcher): real-time search and fact-checking, heavily tapping into X platform data streams
- Benjamin (Logic Expert): step-by-step reasoning, mathematical computation, code generation and verification
- Lucas (Devil's Advocate): specifically finds flaws — identifies bias, questions conclusions, prevents overconfidence
Whenever a sufficiently complex problem arises, the four agents follow a pipeline: task decomposition → parallel analysis (shared KV cache) → multi-round debate → final synthesis.
The debate is not a performance. Lucas exists precisely to break the confirmation bias that Grok and Benjamin might develop.
The hallucination number is striking
The hallucination rate dropped from 12% on Grok 4.1 to 4.2%, a 65% reduction.
At 4%, that is quite low among current large models. OpenAI and Anthropic's flagship models sit around 5-8%.
Other metrics:
| Metric | Grok 4.20 |
|---|---|
| IFBench (instruction following) | 83%, ranked first |
| Inference speed | 220.5 tokens/sec |
| SWE-bench (coding) | 75% |
| Intelligence Index | 8th, score 48 |
| Context window | 2 million tokens |
At 75% on coding, it still trails Claude Opus 4.6's 80.8%. Overall ranking is 8th; GPT-5.4 scores 57, a notable gap. So Grok 4.20 does not claim the "smartest" crown, but it is competitive on "most reliable in what it says."
Why a shared backbone beats multi-model collaboration
Multi-agent frameworks are not new — LangGraph, AutoGen, and others have been around for a while. But xAI's approach has a key difference: instead of having multiple independent models collaborate, it runs multiple roles on the same backbone.
The benefits:
- Shared KV cache, much lower latency
- No need to pass large contexts between models, reducing information loss
- All agents share the same underlying understanding of the same input
In other words, previous "multi-agent" setups were like several people solving a problem together; Grok 4.20 is one person running four modes of thinking simultaneously.
Infrastructure
The system runs on xAI's Colossus supercomputer in Memphis: more than 300,000 GPUs, 2 GW power consumption, 194 PB/s memory bandwidth. The 2-million-token context window can hold a large codebase plus years of documentation.
One detail: xAI has not disclosed whether the number of debate rounds is fixed or adaptive — that part of the engineering remains a black box.
Pricing and access
- API: $2/M input, $6/M output
- Consumers: SuperGrok subscription (approx. $30/month) or X Premium+
The price is higher than Claude Opus 4.6 and slightly lower than GPT-5.4.
This direction is worth studying
The sharp drop in hallucination rate shows that the mechanism of "self-questioning" is engineering-effective. This may be a more worthwhile direction than simply stacking more parameters.
But several questions remain unanswered: What is the control logic for debate rounds? Could Lucas's questioning lead to excessive conservatism? xAI has not disclosed these.
What can be said about the architecture is that this is the first known commercial product to embed four-agent debate into a unified backbone with shared KV cache — if the approach proves effective, it is only a matter of time before other major players follow.
Sources: Inside Grok 4.20: How Four Agents on One Backbone Beat Separate Models (Medium, Engineer at Heart); CocoLoop, Grok 4.20 Beta Launch: 4-Agent AI System Launches (adwaitx.com); HOW THE XAI GROK 4.20 AGENTS WORK (NextBigFuture.com); Master the 5 Core Capabilities of Grok 4.20 (Apiyi.com)