# Grok 4.6

SpaceXAI's frontier model for work that runs long. Grok 4.6 pairs a 500K tokens context window with four reasoning levels, so a single model covers quick agent steps and deep multi-hour tasks.

- **Model ID:** `spacexai/grok-4.6`
- **Type:** chat
- **Providers:** xai
- **Context window:** 500,000
- **Maximum output tokens:** 500,000
- **Pricing:** $2/1M input tokens, $6/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/grok-4.6

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'spacexai/grok-4.6',
  prompt: 'Why is the sky blue?'
})
```

## About

Grok 4.6 was released August 12, 2026 as SpaceXAI's frontier reasoning model for long-running agents, coding, and knowledge work. It refines Grok 4.5 rather than replacing the foundation underneath it, so the shape of the model is familiar and the gains show up in sustained agentic work. Grok 4.6 accepts text and image input, returns text, and works within a context window of 500K tokens. SpaceXAI lists a knowledge cutoff of February 1, 2026.

Reasoning depth is a per-request setting, and Grok 4.6 adds a fourth level. Alongside low, medium, and high, an `xhigh` setting pushes the model further on problems that reward more deliberation. High remains the default. On longer trajectories Grok 4.6 checks its own work before moving on, which is the behaviour that matters most when an agent chains dozens of steps without a human in the loop.

Grok 4.6 leads on GDPval-AA v2 and AA-Briefcase, and scores 61 on the Artificial Analysis Intelligence Index. Coding results are more mixed: it trails the strongest models on DeepSWE v1.1 and Terminal-Bench v3.0, so benchmark it against your own repository before switching a coding agent over.

Pricing is tiered by prompt length, with a higher rate once a request crosses 200K tokens. See the pricing panel on this page for current rates before you send very long prompts.

You can integrate Grok 4.6 through AI SDK, Chat Completions API, Responses API, Messages API, or other API formats, from TypeScript or Python. Routing rules move traffic from another Grok model to Grok 4.6 without changing application code.

## What to consider

Reasoning stays on for every Grok 4.6 request, so responses carry thinking tokens even at the low level. The new `xhigh` level spends more of them than high, which is the point, but it makes per-request output cost harder to predict. Set the level per request rather than leaving everything at the default.

Prompt length changes the rate. Requests at or above 200K tokens bill at a higher tier, and that tier applies to the whole request rather than only the tokens past the threshold, so a prompt that drifts just over the line costs more than one that stays just under it.

Coding is the weaker column on the launch benchmarks. If your workload is mostly software engineering rather than general knowledge work, compare Grok 4.6 against the alternatives below on your own tasks. The February 1, 2026 knowledge cutoff also means Grok 4.6 needs web search or retrieval for anything more recent.

## When to use

### Best For

- **Long-running agents** on multi-hour trajectories where the model verifies its own work between steps
- **Extended reasoning depth** using the fourth `xhigh` level on problems that reward more deliberation
- **Knowledge work** across research, analysis, and briefing tasks where it leads on GDPval-AA v2
- **Large context tasks** that fill a 500K tokens window with long documents and full session history
- **Vision-assisted work** pairing screenshots, diagrams, and design references with text prompts

### Consider Alternatives When

- **Heavy coding workloads** dominate, where Grok Code Fast 1 is tuned for quick code edits at lower cost
- **Reasoning-free responses** are enough, where Grok 4.1 Fast Non-Reasoning answers without spending thinking tokens
- **Short prompt volume** is the pattern, where Grok 4.5 avoids the long-prompt pricing tier
- **Multi-agent orchestration** fits the problem, where Grok 4.20 Multi-Agent spreads one request across collaborating agents

## Best for

- **Long-Running Agents:** Multi-hour trajectories where the model verifies its own work between steps
- **Extended Reasoning Depth:** A fourth `xhigh` level for problems that reward more deliberation
- **Knowledge Work:** Research, analysis, and briefing tasks where it leads on GDPval-AA v2
- **Large Context Tasks:** A 500K tokens window that holds long documents and full session history
- **Vision-Assisted Work:** Screenshots, diagrams, and design references alongside text prompts

## Consider alternatives

- **Heavy Coding Workloads:** Grok Code Fast 1 is tuned for quick code edits at lower cost
- **Reasoning-Free Responses:** Grok 4.1 Fast Non-Reasoning answers without spending thinking tokens
- **Short Prompt Volume:** Grok 4.5 avoids the long-prompt pricing tier on smaller requests
- **Multi-Agent Orchestration:** Grok 4.20 Multi-Agent spreads one request across collaborating agents

## Frequently asked questions

### What is Grok 4.6 built for?

Long-running agents, coding, and knowledge work. Grok 4.6 is tuned for trajectories that continue across many steps, where it verifies its own work before moving on.

### How does Grok 4.6 differ from Grok 4.5?

It refines Grok 4.5 rather than replacing the foundation underneath it. The visible changes are a fourth `xhigh` reasoning level, stronger results on knowledge work benchmarks, and more self-checking on long trajectories.

### What reasoning levels does Grok 4.6 support?

Four: low, medium, high, and xhigh. High is the default. Use low for latency-sensitive agent steps, high for most engineering work, and xhigh for problems that justify the extra thinking tokens.

### What is the context window for Grok 4.6?

The context window is 500K tokens, with up to 500K tokens per response.

### Why does a long prompt cost more on Grok 4.6?

Pricing is tiered by prompt length, and requests at or above 200K tokens bill at the higher tier. The tier applies to every token in the request, not only the tokens past the threshold. See the pricing panel on this page for current rates.

### Is Grok 4.6 a good fit for a coding agent?

It depends on the workload. Grok 4.6 leads on knowledge work benchmarks but trails the strongest models on DeepSWE v1.1 and Terminal-Bench v3.0. Benchmark it against your own repository before switching a coding agent over.

### What is the knowledge cutoff for Grok 4.6?

February 1, 2026. Pair Grok 4.6 with web search or retrieval for facts that changed after that date.

### Can I move traffic to Grok 4.6 without changing my code?

Yes. Add an AI Gateway routing rule that rewrites requests from another Grok model to `spacexai/grok-4.6`. Your application keeps sending the old model identifier.

### Does Grok 4.6 support Zero Data Retention?

Yes, Zero Data Retention is available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

## Links

- [Model page](https://vercel.com/ai-gateway/models/grok-4.6)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)


If you encounter incorrect information, can’t find what you need, need a capability Vercel doesn’t currently offer, or have any other feedback to report, please do it by following the instructions at https://vercel.com/feedback.md.