Updated August 13, 2026.
On August 12, SpaceXAI released Grok 4.6 — a new frontier model for programming, long-horizon agent tasks, and knowledge work. It includes a 500,000-token context window, text and image input, four levels of reasoning, and built-in tools. API pricing starts at $2 per 1 million input tokens and $6 per 1 million output tokens.
In brief: the main change in Grok 4.6 is not a new chat interface, but better performance across long work trajectories: the model was trained on agentic RL tasks, code, CAD, web development, and knowledge work. According to SpaceXAI’s published tests, it is noticeably stronger than Grok 4.5, but this still needs to be validated in your own harness.
This material is intended for developers, CTOs, and AI product owners. It covers the API model grok-4.6, not consumer SuperGrok pricing and not separate Imagine or Voice products.
Contents
- What’s new in Grok 4.6
- Specifications and API
- What the tests show
- How much Grok 4.6 costs
- How to connect the model
- How Grok 4.6 differs from 4.5
- What tasks it is suitable for
- Risks and limitations
- FAQ
- How AI Dawn helps you evaluate Grok on a business task
- Bottom line
What’s new in Grok 4.6
In the official announcement SpaceXAI emphasizes long agentic scenarios: topic research, working with a codebase, information analysis, and building interactive applications. The model went through a longer additional training run than 4.5, after which the SFT trajectories were regenerated from Grok 4.5 for different reasoning efforts and the agent harness.
The developer also reports expanded RL on knowledge work, general and specialized code, kernel optimization, web development, and CAD. On long tasks, the team observed more self-checking, and in visual and interactive projects, a stronger first pass. These are the vendor’s observations, not an independent AI Dawn test.
Specifications and API
| Parameter | Grok 4.6 |
|---|---|
| Model ID | grok-4.6 |
| Context | 500,000 tokens |
| Knowledge cutoff | February 1, 2026 |
| Input | text and images |
| Output | text; no separate limit is specified in the documentation |
| Reasoning | low, medium, high by default, xhigh |
| API | Responses API and Chat Completions |
| Tools | function calling, web search, X search, code execution |
| Availability | xAI API, Grok Build, Cursor, OpenRouter, Vercel, and Cloudflare |
A knowledge cutoff does not mean the model only knows older events: web search and X search can bring in current data during a request. But without a tool, information after February 1, 2026 should not be considered built into the weights.
What the tests show
SpaceXAI publishes a comparison of Grok 4.6 High with Grok 4.5 High and competitors. Some comparisons are taken from self-reported results and public leaderboards, so the table helps narrow down candidates but does not replace a single independent run.
| Benchmark | Grok 4.6 High | Grok 4.5 High | Change |
|---|---|---|---|
| AA Intelligence Index | 61 | 56 | +5 points |
| GDPVal-AA v2 | 1753 | 1526 | +227 |
| CursorBench v3.2 | 69.9% | 66.7% | +3.2 pp |
| DeepSWE v1.1 | 65.9% | 54.0% | +11.9 pp |
| FrontierCode v1.1 Extended | 61.3% | 56.6% | +4.7 pp |
| Terminal-Bench v3.0 | 26.0% | 15.7% | +10.3 percentage points |
In the AA Intelligence Index, Grok 4.6 scored 61 — the same as GPT-5.6 Sol Max in the announcement table. That does not mean the models are equal on every task: the index combines nine tests, and individual model profiles differ.
How much does Grok 4.6 cost?
The standard API rate starts at $2 per 1M input tokens and $6 per 1M output tokens. The fast option costs twice as much: $4 and $12 respectively. Exact cached input pricing and rate limits are shown on the team detail page and may depend on the account.
Example: 100,000 input and 20,000 output tokens for the standard option cost 0.1 × $2 + 0.02 × $6 = $0.32. This calculation excludes web/X search, retries, caching, and result verification.
For a long-running agent, calculate the cost of the accepted task, not the call. That includes reasoning, replayed history, tool results, retries, and human review.
How to connect the model
from openai import OpenAI
import os
client = OpenAI(
api_key=os.environ["XAI_API_KEY"],
base_url="https://api.x.ai/v1",
)
response = client.responses.create(
model="grok-4.6",
input="Find the bug, fix it, and suggest a regression test",
reasoning={"effort": "high"},
prompt_cache_key="project-auth-migration",
)
print(response.output_text)SpaceXAI recommends passing prompt_cache_key in the Responses API or the x-grok-conv-id header in Chat Completions. This routes requests from the same conversation to the same server and increases the likelihood of a cache hit; without a key, a long history is often billed as cold input.
For long-running workflows, use context compaction. Do not let the model automatically trim critical requirements, decision logs, and security policy: it is better to keep those blocks separately and restore them explicitly.
How Grok 4.6 differs from 4.5
| Area | Grok 4.5 | Grok 4.6 |
|---|---|---|
| Main focus | coding, agents, knowledge work | longer agents, visual and interactive projects |
| Context in current docs | previous-generation model | 500K tokens |
| Reasoning | intelligent/efficient mode | low, medium, high, xhigh |
| API pricing | $2 / $6 per 1M | starting at $2 / $6; fast is twice as expensive |
| Published evals | baseline | growth across all shown xAI tests |
Migration makes sense if long tasks in 4.5 get cut off, the agent loses the objective after a series of tools, or the first pass through an app requires too many manual iterations. For short classification tasks, the update may not deliver an economic benefit.
What tasks is it suitable for?
- modifying multiple related modules and running tests independently;
- researching a topic with web/X search and a summary report;
- prototyping an interactive application;
- analyzing large documents and images;
- long agentic workflows with function calling and code execution.
Do not give an agent irreversible permissions by default. Payments, data deletion, publishing, and production deployment should require separate confirmation.
Risks and limitations
First, announcement results are not your benchmark. The harness, reasoning effort, tools, and time limit all change the answer. Second, 500K tokens is capacity, not a guarantee that every detail will be retained. Third, web/X search brings in external untrusted content and creates a prompt injection risk.
A pilot should include 50–100 real anonymized tasks, the same permissions, a predefined success criterion, p50/p95 latency, tokens, number of retries, and the share of results accepted by a human. This is a practical recommendation, not a statistical guarantee.
FAQ
When was Grok 4.6 released?
SpaceXAI released the model on August 12, 2026. The official announcement and API documentation were updated the same day.
What context window does Grok 4.6 have?
500,000 tokens. For long agent loops, the documentation recommends prompt caching and context compaction.
How much does Grok 4.6 cost?
The standard API is $2 per 1M input tokens and $6 per 1M output tokens. The fast option costs twice as much.
Does Grok 4.6 support images?
Yes, the official model card indicates text and image input, while the output is text-only.
Does it support internet and X search?
Yes. Web search and X search are available as server-side tools alongside code execution and function calling.
How AI Dawn helps validate Grok on a business use case
AI Dawn can prepare a test set and acceptance criteria, design the RAG and data pipeline, integrate the agent with enterprise systems, and set up confirmation for risky actions. For a closed environment, the team can compare an external API with an on-premise LLM based on data requirements and total cost of ownership.
The first step is to choose one process, establish the baseline, data sources, constraints, and acceptance criteria. Discuss the project.
Bottom line
Grok 4.6 is an evolution of 4.5 with a clear focus on long-running agents, code, and visual-interactive projects. The model got a 500K context window, xhigh reasoning, images, and the full set of server-side tools at the same starting price of $2/$6.
Published tests show improvement, but the migration decision should be based on your own tasks. Run 4.5 and 4.6 in the same harness, measure result acceptance, latency, and total cost — and update the route only where the difference is reproducible.