Grok 4.20

Grok 4.20 at a glance

Grok 4.20 is the model our chat uses unless you pick another one. It is built for speed: in our setup it replies straight away, without spending time on a hidden thinking step, so the first words of a short answer usually appear within a second or two. It understands images as well as text, writes clearly in many languages and handles everyday work such as summaries, emails, explanations, translations and simple code without fuss.

• Maker: xAI, now part of SpaceX and branded SpaceXAI / SpaceXSI.
• Released: as a beta in the Grok apps in February 2026; on the xAI API on March 10, 2026.
• Input: text and images. Output: text.
• Context window: 2,000,000 tokens on OpenRouter; xAI's own docs list 1,000,000.
• In our chat: the default model, with reasoning switched off.

Where Grok 4.20 came from

Grok 4.20 followed Grok 4.1, which xAI shipped in November 2025. It was expected earlier, but reached grok.com and the phone apps as a beta in mid-February 2026. That consumer version worked as a small team: four agents, named Grok, Harper, Benjamin and Lucas in press coverage, looked at a question in parallel before a lead agent wrote the final answer.

On March 10, 2026, xAI put the model on its API in three forms. Two are single models, a reasoning version and a non-reasoning one, both under the grok-4.20 name. The third, Grok 4.20 Multi-Agent, runs four or sixteen agents together for deep research, according to xAI's multi-agent documentation. xAI's model page describes Grok 4.20 as having industry-leading speed and strong tool calling, and claims a very low rate of made-up facts. Those are the maker's own claims; we have not measured them.

The model did not stay the newest for long. Grok 4.3 arrived at the end of April and Grok 4.5, 4.6 and 4.7 followed over the summer. Grok 4.20 kept its place as the quick, low-cost option.

How Grok 4.20 works in our chat

We run the single-model version through OpenRouter with reasoning turned off. That choice matters. Without the thinking step, Grok 4.20 starts writing at once, which feels natural for conversation and keeps our running costs low enough to offer 30 free messages a day. The trade-off is that it does not double-check itself on tricky logic the way Grok 4.3 Thinking does.

You can attach one image per message, such as a photo, a chart or a screenshot, and ask questions about it. Our chat sends the recent part of the conversation with each message rather than the full multi-million-token window, so for a very long document, paste the section you care about.

Grok 4.20 in our chat cannot browse the web or read X. It answers from what it learned in training. The multi-agent research version is not part of our chat.

That makes our version different from the one people met on grok.com in February. There, the beta combined several agents with live search, which is why its answers could quote fresh posts and links. Here you get the plain model: fast, consistent and cheap enough to offer free every day, but working only from its own knowledge.

What Grok 4.20 is good at

• Quick writing jobs. Replies, short posts, product descriptions and polite refusals. Try: decline a meeting politely.
• Summaries. Paste notes or an article and ask for the key points and next steps: summarize meeting notes.
• Plain explanations. Concepts from school, work or the news, at the level you ask for: compound interest for a 12-year-old.
• Images. Describe a photo, read a sign, explain a chart or spot a problem in a screenshot: ask about a photo.
• Translation and rewriting between many languages, including tone changes such as formal to casual.
• Small coding tasks like explaining an error message or writing a short function.

Where it falls short

• Multi-step maths and logic. Without reasoning, it can jump to an answer that looks right and is not. Use Grok 4.3 Thinking for puzzles, proofs and calculations with several steps.
• Large, tricky code changes. Grok 4.7 is much stronger on real software engineering.
• Recent events. It has no search in our chat, so it may not know about things that happened after its training.
• Confident mistakes. Like any language model, it can state wrong facts fluently. Check anything important.

Getting better answers from Grok 4.20

• Be specific about the output. "Three bullet points, under 60 words" beats "summarize".
• Give it the material. Paste the email, the paragraph or the error message instead of describing it.
• Ask for a check when it matters. "List any assumptions you made" helps catch shortcuts.
• Switch models mid-chat. If an answer needs more care, pick Grok 4.3 Thinking from the model menu and ask again.
• Use a tool page for routine jobs: our summarizer and translator are tuned for one-shot results.

Our comparison of Grok 4.20 and Grok 4.3 explains when the slower model is worth it.

Chat with Grok 4.20, free and without an account.

FAQ

Why is it called Grok 4.20?

xAI has not explained the number in its documentation. It sits between Grok 4.1 (November 2025) and Grok 4.3 (April 2026) in the line-up.

Does Grok 4.20 think before answering?

xAI offers both a reasoning and a non-reasoning version. In our chat, reasoning is switched off so answers start right away.

Can Grok 4.20 read images?

Yes. Attach a photo or screenshot to your message in our chat and ask about it. It cannot create images.

Is Grok 4.20 free on Grok.Online?

Yes. It counts toward the 30 free messages each visitor gets per day, with no separate limit.

What does Grok 4.20 cost on the API?

$1.25 per million input tokens and $2.50 per million output tokens (OpenRouter list price, Oct 2026), the same as xAI's base rate for prompts under 200,000 tokens.