Grok 4.20 and Grok 4.3

What Grok 4.20 is

Grok 4.20 is a family of large language models from xAI, reachable both through the Grok apps and through xAI's API. The API exposes it as three variants released together: grok-4.20-0309-reasoning, grok-4.20-0309-non-reasoning and grok-4.20-multi-agent-0309^[3]. The names combine a version number with a date code, so a caller can tell from the model name which build it is pointing at.

The reasoning and non-reasoning variants differ in how they spend computation. A reasoning model works through intermediate steps before it produces an answer, which costs more tokens and more time but helps on problems that need several steps of deduction; a non-reasoning model replies directly, which is cheaper and faster for routine requests. Shipping both under one release lets a developer send easy calls to the cheap variant and hard ones to the reasoning variant without switching model families.

All of them share a context window of 1 million tokens^[4]. The context window is the amount of material a model can consider in a single request — the prompt, any attached documents or images, and the answer generated so far. A million tokens is on the order of hundreds of thousands of words, enough to hold a long book or a sizeable code base in one call. The models take text and images as input and return text, and they support function calling (letting the model ask for a named tool to be run), structured outputs (forcing the reply into a fixed shape such as JSON) and reasoning^[4].

Timeline

Grok 4.20 arrived in the Grok apps in February 2026^[1]. xAI's API release notes list Grok 4.20 and Grok 4.20 Multi-agent as live from 10 March 2026^[2]. Grok 4.3 followed in April 2026, keeping the same 1-million-token context and the same API prices^[1][3].

The multi-agent variant

In grok-4.20-multi-agent-0309, several agents work in parallel on deep-research tasks^[4]. The idea behind this arrangement is that a broad research question can be split into threads that are pursued at the same time and then brought back together, instead of being worked through one step after another. The documentation lists aliases including grok-4.20-multi-agent-latest and grok-4.20-multi-agent-beta-0309, served from us-east-1 and us-west-2^[4]. Its documented rate limits were 9 requests per second and 2.5 million tokens per minute, and it supported the Batch API^[4].

Pricing

For prompts under 200,000 tokens, the API price was $1.25 per million input tokens, $0.20 per million cached input tokens and $2.50 per million output tokens; above that length the rates doubled, and requests sent through the Batch API got 20% off^[4]. The tiers follow the cost of serving: a very long prompt occupies more of the model's working memory for longer, so it is charged at a higher rate, while cached input is cheaper because a repeated prefix does not have to be processed again. Batch requests are discounted because they can be queued rather than answered immediately.

Where it sits

Grok 4.5 followed in July 2026, described as a coding and agentic model co-trained with Cursor^[1]. The earlier pair did not disappear when it arrived: Grok 4.20 and Grok 4.3 were both still listed as current API models in November 2026^[3]. For developers, the practical shape of the release was a single 1-million-token family split by how much thinking a request needs — no reasoning, reasoning, or several agents working side by side — with prices that stayed unchanged between the 4.20 and 4.3 generations^[3].