xAI Launches Grok 4.20 and $10 SuperGrok Lite Subscription

On March 26, 2026, xAI released Grok 4.20 and a new $10 SuperGrok Lite subscription. The release introduces a specialized routing architecture designed to lower error rates while pushing API costs down to $2.00 per million tokens. For developers building systems that require verifiable citations, the model’s new internal consensus approach changes how reasoning steps are handled in production.

Adversarial Consensus Architecture

Grok 4.20 achieves a 65% reduction in hallucinations compared to Grok 4.1. The model’s baseline error rate drops from 12% to 4.2%. This accuracy improvement relies on an internal system called Adversarial Consensus.

Every query automatically triggers four specialized internal agents before generating an output. The agents include Grok acting as captain, Harper checking facts, Benjamin handling math and logic, and Lucas providing contrarian or creative angles. These nodes debate and cross-check each other to finalize a response.

If you design multi-agent systems, this native routing shifts the debate mechanism from your application layer directly into the model’s inference process. The model supplements this reasoning with direct access to X Search and general Web Search. This produces inline citations linking to specific platform posts and external URLs.

Context Window and API Economics

The model supports a 2.0 million token context window. It targets a low latency profile of approximately 250 milliseconds per request.

The API pricing sits at $2.00 per million tokens. xAI positions this as up to 64x cheaper than previous frontier reasoning models. At this price point, feeding large codebases or extensive document libraries into the prompt becomes viable for routine processing tasks.

SuperGrok Subscription Tiers

The consumer rollout introduces SuperGrok Lite at $10 per month. The tier targets casual creators who need multimodal tools but do not require enterprise compute. Users gain access to basic AI image generation through the Aurora model and video generation via Grok Imagine.

Video outputs on the Lite tier are restricted to 480p resolution and a maximum duration of six seconds per clip, with an undisclosed daily generation cap. The plan also doubles the chat session limits compared to the free tier and allows users to configure one custom agent in Expert mode.

Subscription Tier	Monthly Price	Video Capability	Agent Access
SuperGrok Lite	$10	480p, 6 seconds	1 custom agent
SuperGrok Heavy	$300	720p, 30 seconds	16-agent reasoning

Infrastructure and Production Context

Grok 4.20 was trained on Colossus, xAI’s 200,000 GPU cluster located in Memphis. The release follows rapid structural changes at the company. In January 2026, xAI raised $20 billion at a $230 billion valuation. Weeks later, SpaceX acquired the organization to integrate AI models across its aerospace and Starlink operations. The aggressive release cadence reflects this newly consolidated compute footprint.

Evaluate Grok 4.20’s API if your current stack relies on external orchestration frameworks to force model consensus. You can likely strip out custom debate prompts and reduce your tool-calling latency by letting the native Adversarial Consensus handle fact-checking internally. Test the inline citation feature against your existing retrieval systems to see if the integrated X Search provides better source attribution for real-time queries.