Grok 4.20

Overview

Freshness note: Model capabilities, limits, and pricing can change quickly. This profile is a point-in-time snapshot last verified on May 8, 2026.

Grok 4.20 is no longer just a “coming soon” placeholder. xAI’s March 10, 2026 release notes now say both Grok 4.20 Beta and Grok 4.20 Multi-agent Beta are available in the xAI Enterprise API. That is still not the same thing as a quiet, long-settled general-availability model page, but it does move 4.20 from rumor or waitlist territory into a real preview surface.

The right way to read this page is “active preview research lane.” xAI’s current models page lists the 4.20 variants with a 2M-token context window and public token pricing, but the same page strongly recommends Grok 4.3 for standard chat, coding, and general API callers. Treat 4.20 as a deliberate beta evaluation, not the default Grok route.

Capabilities

The clearest documented capabilities today are speed, strict instruction following, large context, and multi-agent research. xAI’s current docs and guides also show 4.20 working with server-side tools such as web_search, x_search, code_execution, collections_search, and remote MCP tools.

The multi-agent variant is especially notable because it turns Grok from a single-model assistant into a coordinated research system. That matters for tasks where multiple sub-questions should be explored in parallel and then synthesized by a lead agent.

Technical Details

xAI’s current official docs support these technical anchors:

• 2,000,000-token context window on the models page.
• beta single-agent and multi-agent 4.20 variants in the Enterprise API.
• active beta model names such as grok-4.20-beta-latest-non-reasoning in current batch examples and grok-4.20-multi-agent-beta-0309 in multi-agent documentation.
• explicit limits such as no logprobs support on Grok 4.20 models.

The big operational point is that 4.20 should still be treated as preview infrastructure. It is more real and more available than the older repo copy suggested, but it is not the retirement replacement for Grok 4, Grok 4.1 Fast, or Grok Code Fast 1. xAI points those migrations at Grok 4.3.

Pricing & Access

xAI’s public models page currently lists the 4.20 chat variants at:

• Input: $1.25 per 1M tokens
• Output: $2.50 per 1M tokens

Operational notes:

• Grok 4.20 remains a preview/enterprise-oriented lane.
• Multi-agent requests inherit both model-token costs and server-side tool costs.
• beta multi-agent runs can be materially more expensive than single-agent requests.
• teams should verify access and billing directly in the current xAI console and docs before production planning.

Access options:

• xAI Enterprise API
• beta multi-agent research interface
• batch and agent-tool workflows where enabled

xAI’s broader API remains OpenAI-compatible, which lowers integration friction once access is available.

Best Use Cases

Grok 4.20 fits best when the real job is research orchestration:

• Deep research tasks where several sub-questions should be explored in parallel.
• Multi-source synthesis that benefits from web, X, document, and code tools in one loop.
• Experimental research agents where a team can tolerate beta-model volatility.
• Internal evaluation of xAI’s multi-agent direction before wider deployment.

It is less suitable for budget-sensitive production traffic, strict SLAs, or cases where teams need a fully documented stable contract today.

Comparisons

• Grok 4.3 (xAI): Recommended default for general, coding, reasoning, and non-reasoning API work.
• Grok 4 (xAI): Retiring older premium Grok path; migrate new work to 4.3 instead.
• Claude Opus 4.7 (Anthropic): Stronger choice when governance, documentation depth, and enterprise stability matter more than experimental multi-agent breadth.