Grok 4.3 Thinking

Grok 4.3 at a glance

Grok 4.3 is a reasoning model. Before it writes a reply, it works through the problem privately: it breaks the question into steps, tries an approach, checks it and only then answers. In our chat we call it Grok 4.3 Thinking so you know what to expect, which is a short pause and then a more careful answer. It is the model to reach for when a question has a right and a wrong answer and getting there takes several steps.

• Maker: xAI, now part of SpaceX and branded SpaceXAI / SpaceXSI.
• Released: April 30, 2026 on the xAI API and OpenRouter, after a beta for subscribers earlier that month.
• Input: text and images. Output: text.
• Context window: 1,000,000 tokens.
• Knowledge cut-off: December 2025, according to Oracle's documentation for the model.
• In our chat: low reasoning effort, with the thinking hidden and only the final answer shown.

What is new in Grok 4.3

Grok 4.3 arrived about seven weeks after Grok 4.20 reached the API. xAI's model page describes it as a fast, reliable model with strong tool calling and instruction following, and Oracle, which also hosts it, calls it a reasoning model for complex, accuracy-critical work with an improved architecture compared with Grok 4.20.

Three things stand out for everyday users:

• Adjustable thinking. Developers can set the reasoning effort from none up to xhigh. The default is low, which is also what we use.
• A very long context. One million tokens is enough for several books' worth of text in a single request, although our chat sends only the recent part of a conversation.
• The same price as Grok 4.20. Both are listed at $1.25 for a million input tokens and $2.50 for a million output tokens (OpenRouter list price, Oct 2026), so Grok 4.3 is not more expensive per token. It does use extra output tokens for its thinking, so each answer costs a little more.

How Grok 4.3 Thinking works in our chat

When you pick Grok 4.3 Thinking from the model menu, the chat shows a "Thinking…" line while the model reasons. On a short question this takes a few seconds; on a hard one, longer. Then the answer streams in as usual. You do not see the hidden working-out, only the result, which keeps replies readable.

It reads images too, so you can photograph a worksheet, a diagram or a table and ask it to work through the problem. It counts toward the normal 30 free messages a day, with no separate cap.

Like every model on our site, it cannot browse the web or read X, and it does not create images. Its knowledge stops at its training cut-off.

Where the thinking pays off

• Maths and word problems, with steps shown so you can follow them: a timetable question.
• Logic and arguments. Finding the gap in reasoning, checking a conclusion, comparing options fairly: spot the flaw.
• Planning. Schedules, budgets and step-by-step plans that have to add up: a five-day study plan.
• Decisions with trade-offs, such as renting versus buying, laid out in a table.
• Debugging, where tracing what the code actually does matters more than writing new code.
• Careful reading of contracts, instructions or rules, when you need to know exactly what a passage says.

A small example

Ask a fast model: "A bat and a ball cost 1.10 together, and the bat costs 1.00 more than the ball. How much is the ball?" A quick reply often says 0.10, because that is the number that jumps out. A thinking model is more likely to set up the two amounts, test the first guess (a 0.10 ball would make the bat 1.10 and the total 1.20), notice the mismatch and arrive at 0.05. That habit of testing an answer before giving it is exactly what you are paying for with the extra seconds, and it matters most on homework, budgets and anything with several conditions.

When another model is the better choice

• Simple, quick tasks. For a short email or a translation, the pause adds nothing. Grok 4.20 is faster.
• Hard engineering and long agent-style tasks. Grok 4.7 is newer, larger and stronger at complex code, though limited to 10 messages a day in our chat.
• Events after December 2025. Grok 4.7's training goes up to May 2026, so it knows more recent background, but no model here can check today's news.

Thinking makes mistakes less likely, not impossible. Check numbers and facts that matter, and ask the model to show its steps so you can follow the logic yourself. Our guide to reasoning models explains why they sometimes still slip.

Prompting tips for a thinking model

• State the goal and the constraints ("budget under 500 euros, two people, three days") rather than steering every step. Reasoning models plan better when they know the target.
• Ask for the steps when you need to learn or verify, and for "just the answer" when you do not.
• Give all the data in one message. Numbers spread across several messages are easier to miss.
• Ask it to check its result another way, for example by estimating or working backwards.

Chat with Grok 4.3 Thinking, free and without signing up.

FAQ

What does "Thinking" mean in Grok 4.3 Thinking?

The model reasons privately before it replies. We run it at low effort and show only the final answer.

Is Grok 4.3 better than Grok 4.20?

For maths, logic and planning, usually yes. For quick everyday writing, Grok 4.20 is faster and just as useful.

Why does Grok 4.3 take longer to answer?

It spends a few seconds working through the problem first. The harder the question, the longer that step can take.

Can Grok 4.3 read images?

Yes. Attach a photo, chart or screenshot and ask about it. It cannot generate images.

How many Grok 4.3 messages can I send?

It shares the 30 free messages per day with the other models and tools on our site.