Grok 4.3 Guide: The Cheap 1M-Context Tier of the Grok Family

Every model family ends up with a tier that does most of the actual work: cheap enough to run on everything, capable enough that you rarely need to escalate. In the current Grok line-up that tier is Grok 4.3. It is not the flagship, and xAI does not pitch it as one. Its model page calls it a "fast, reliable model with strong tool calling and instruction following capabilities".

A note on names: xAI, the company behind Grok, was renamed SpaceXAI in July 2026 after SpaceX acquired it. This guide uses the xAI name throughout.

What makes it interesting is the combination. It keeps the 1M-token context window that the newer Grok 4.5, 4.6 and 4.7 cut to 500K, it is listed at less than half their output price, and it has the finest-grained reasoning control in the family.

Where Grok 4.3 sits

The Grok API currently has two price tiers. Grok 4.20 and Grok 4.3 are listed at $1.25 per million input tokens, $0.20 cached and $2.50 output. Grok 4.5, 4.6 and 4.7 are listed at $2.00 input and $6.00 output. OpenRouter shows the same base rates for Grok 4.3 at the time of writing, and lists the model from 30 April 2026.

On the spec sheet:

• Context: 1,000,000 tokens.
• Input: text and images; output is text.
• Capabilities: function calling, structured outputs and reasoning.
• Rate limits: 37 requests per second and 10 million tokens per minute at the time of writing.
• Regions: us-east-1, us-west-2 and eu-west-1. At the time of writing, the 4.5, 4.6 and 4.7 model pages list only US regions, so this matters if data residency is a requirement. Check the current page before relying on it.

xAI has not published an architecture, parameter count or benchmark table for Grok 4.3 that we could find, and there are no open weights. The independent reference point is Artificial Analysis, which scored it 53 on its Intelligence Index at launch, four points above the Grok 4.20 version it had tested. Treat that as a coarse ranking and judge it on your own tasks.

Five effort levels, and a low default

Grok 4.3 accepts reasoning_effort values of none, low, medium, high and xhigh. The default is low.

Both details are deliberate and both are easy to miss. none means you can switch reasoning off on the same model, which Grok 4.20 handles with a separate endpoint and the 4.5-and-later models do not offer at all. And a low default means that, out of the box, Grok 4.3 spends little on hidden reasoning. The newer models default to high.

That makes Grok 4.3 easy to route within. Use none for extraction and reformatting, low for ordinary tool use, and escalate effort per request only where a task needs multi-step planning. Reasoning tokens are billed as output, so effort is a direct cost lever. The token cost of reasoning models explains how quickly hidden output adds up.

The batch endpoint

Grok 4.3 supports xAI's Batch API at a 20% discount to standard rates. On OpenRouter the same model appears as x-ai/grok-4.3:batch, listed at $1.00 input and $2.00 output per million tokens at the time of writing. It is the same model on an asynchronous endpoint: you submit work, results arrive later, and you pay less.

The newer Grok 4.5, 4.6 and 4.7 pages all list batch as not supported. If you have a large offline job, such as back-filling summaries, classifying an archive or running an evaluation suite, Grok 4.3 on batch is, alongside Grok 4.20, the cheapest way to run it on a Grok model with a 1M window. Batch API savings math covers when the latency trade is worth it.

1M context, priced in two tiers

The 1M window is real, but it is not priced flat. Once a prompt reaches 200K tokens, xAI bills every token in that request at double the base rate: $2.50 input, $0.40 cached and $5.00 output.

In practice that means three bands of use:

• Under 200K: the base rate, shared with Grok 4.20 and well below the 4.5 tier.
• 200K to 500K: still possible on Grok 4.5 and later too, where the long-context rates are $4 input and $12 output. Grok 4.3 remains the cheaper of the two.
• 500K to 1M: only Grok 4.3 and Grok 4.20 in the current line-up accept prompts this long at all.

Large windows also degrade before they fill, so test recall at the lengths you plan to use: plant a few facts deep in a long prompt and check the model can find them.

Grok 4.3 or Grok 4.5 and later

The step up to Grok 4.5 costs 60% more on input and 140% more on output at list price, halves the context, and drops batch support. What it buys is a model xAI explicitly trained for agentic coding, with a vendor benchmark table on coding tasks that Grok 4.3 never got.

A practical rule:

• Stay on Grok 4.3 for high-volume tool use, document processing, anything over 500K tokens, anything that can run on batch, and workloads that need an EU region.
• Move up for long agentic coding sessions where a failed run costs more than the token difference.
• Route between them when your traffic mixes both. Cheap model first, escalate on failure, is usually the right answer. Model routing and fallbacks covers how to set that up.

How to test it

Pick fifteen tasks from your real workload. Run each at none, low and high effort, and record output tokens and whether the answer is usable. Most teams find a large share of tasks where the cheaper settings are indistinguishable, and that share is where Grok 4.3 pays for itself. Then run the tasks that failed at high on a 4.5-tier model, and you have a measured escalation rule instead of a guess.