⏱️ Lectura: 10 min
xAI released Grok 4.6 on August 12, 2026, and for the first time put it on the same level as GPT-5.6 Sol Max on the Artificial Analysis composite index: 61 points each. The model targets a very specific problem: sustaining long agentic tasks, from researching a topic to building a complete application across multiple iterations, without losing track.
📑 En este artículo
It’s the fourth update to the Grok family in less than a year, and the first that xAI explicitly describes as geared toward more ambitious visual and interactive work, not just solving isolated coding problems.
TL;DR
- xAI released Grok 4.6 on August 12, 2026, the direct successor to Grok 4.5.
- It ties GPT-5.6 Sol Max on the AA Intelligence Index: 61 points each, below Fable 5 Max (62).
- On CursorBench v3.2 it scores 69.9%, versus 66.7% for Grok 4.5 High.
- On Terminal-Bench v3.0 it gets just 26%, far below the 34.6% of GPT-5.6 Sol Max.
- Price: $2 per million input tokens and $6 per million output tokens; the fast variant costs double.
- Available today in Cursor, Grok Build, the xAI API, OpenRouter, Vercel, and Cloudflare.
- Double included usage during the first week in Grok Build and Cursor.
- Trained with an SFT stage regenerated by Grok 4.5 across STEM domains, software engineering, and knowledge work.
Introducing Grok 4.6
Grok 4.6 builds on Grok 4.5 with a particular focus on long-running agents and more ambitious interactive, visual work, according to xAI’s official announcement. The company describes it as a model capable of sticking with multi-step tasks: researching a topic, analyzing information, working across an entire codebase, or turning an idea into a working application.
For a developer in Latin America, the practical question isn’t whether the model is intelligent in the abstract, but whether it’s worth migrating an agentic workflow (for example, a code review pipeline or a research agent) from another provider to Grok 4.6. The numbers xAI published, compared against direct rivals, offer a first answer.
What Happened with Grok 4.6
The launch included immediate availability in Cursor and in Grok Build, xAI’s own tool for building applications with agents. During the first week, xAI is offering double the included usage on both platforms so developers can try the model without exhausting their usual quota.
The model also became available in the xAI API and on partner platforms like OpenRouter, Vercel, and Cloudflare. The list price is $2 per million input tokens and $6 per million output tokens; there’s also a fast variant that costs double per token.
On benchmarks, Grok 4.6 reaches 61 points on the AA Intelligence Index, the same score as GPT-5.6 Sol Max and just one point behind Fable 5 Max (62). The index combines nine different benchmarks, so a tie in the total doesn’t mean a tie on every individual benchmark, something that becomes clear as soon as you open the full table.
Context and History of the Grok Family
Grok 4.5 launched months earlier with an AA Intelligence Index of 56, eight points below the current version. The difference between the two versions doesn’t come from a new architecture but from different training: xAI describes a longer supplementary training stage than the one used for Grok 4.5, with model-generated curated data for reasoning and advanced technical concepts, plus high-quality engineering data and an improved optimizer.
That foundation was later used to regenerate the supervised fine-tuning (SFT) trajectories. xAI used Grok 4.5 itself to produce those trajectories across different reasoning effort levels, different agent harnesses, and different domains (STEM, software engineering, knowledge work), filtering out problematic traces with model-based checks before using them for training.
The result was then refined with reinforcement learning (RL) across a broad range of agentic tasks: knowledge work, general programming, and domain-specific environments like kernel optimization, web development, and computer-aided design (CAD). This is the first time xAI has explicitly mentioned CAD as a training environment for a Grok model, which helps explain the jump in single-shot interface and visual application generation tasks.
Technical Details and Performance of Grok 4.6
xAI published a comparison table with nine benchmarks. Here are the results against Grok 4.5 High, GPT-5.6 Sol Max, and Fable 5 Max:
| Benchmark | Grok 4.6 High | Grok 4.5 High | GPT-5.6 Sol Max | Fable 5 Max |
|---|---|---|---|---|
| AA Intelligence Index | 61 | 56 | 61 | 62 |
| GDPVal-AA v2 | 1753 | 1526 | 1728 | 1741 |
| CursorBench v3.2 | 69.9% | 66.7% | 67.2% | 70.5% |
| DeepSWE v1.1 | 65.9% | 54% | 73% | 70% |
| FrontierCode v1.1 (Extended) | 61.3% | 56.6% | 60.6% | 63.6% |
| APEX-Agents | 57.5% | 47.1% | 56.7% | 59.2% |
| Terminal-Bench v3.0 | 26% | 15.7% | 34.6% | 34.1% |
| APEX-SWE | 56.4% | 53.6% | N/A | 58.8% |
| AA-Briefcase | 1577 | 1313 | 1502 | 1574 |
The pattern is consistent: Grok 4.6 improves on Grok 4.5 across all nine benchmarks, matches or beats GPT-5.6 Sol Max on half of them, but clearly falls behind in two specific cases: DeepSWE v1.1 (65.9% versus 73% for GPT-5.6 Sol Max) and, above all, Terminal-Bench v3.0, where it scores 26% against 34.6% for OpenAI’s model and 34.1% for Fable 5 Max.
⚠️ Heads up: Terminal-Bench v3.0 measures direct interaction with a real terminal. If your workflow depends on the agent executing shell commands autonomously and reliably, Grok 4.6 currently performs below both its direct rivals in that specific scenario.
One detail xAI highlights in the announcement: on long trajectories, the model began showing more self-evaluation and verification, reviewing its own work before moving on to the next step. This is consistent with the stated training goal: RL on agentic tasks where the cost of an error accumulates across many steps, not in a single response.
flowchart TD
A["Product idea"] --> B["Grok 4.6 researches the domain"]
B --> C["Generates structure and first version"]
C --> D["Self-evaluation: reviews its own work"]
D --> E{"Does it meet the goal?"}
E -- "no" --> C
E -- "yes" --> F["Iteration with user feedback"]How to Get Started with Grok 4.6
To try Grok 4.6 today, there are two paths: the Grok Build CLI or the direct API. The CLI installation uses the same script on macOS and Linux; on Windows, the recommended way is to run it inside WSL (Windows Subsystem for Linux), since xAI didn’t publish a native PowerShell installer for this release.
# macOS and Linux (also inside WSL on Windows)
curl -fsSL https://x.ai/cli/install.sh | bash
# verify it installed correctly
grok --versionThat first block just confirms the binary landed in the PATH. The next step is generating a key from the xAI console and calling the model directly via the API, which is compatible with the chat completions format already used by most SDKs:
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["XAI_API_KEY"],
base_url="https://api.x.ai/v1",
)
respuesta = client.chat.completions.create(
model="grok-4.6",
messages=[
{"role": "system", "content": "You are an agent that reviews Python pull requests."},
{"role": "user", "content": "Review this diff and point out concurrency bugs: <diff>"},
],
)
print(respuesta.model)
print(respuesta.choices[0].message.content)</diff>The respuesta.model field returns the exact identifier of the model that responded (for example grok-4.6). It’s the most direct way to confirm your call is actually using the new version and didn’t default back to Grok 4.5 on a router you forgot to update.
💡 Tip: If you integrated Grok through OpenRouter, Vercel, or Cloudflare, check each platform’s model selector: all three added Grok 4.6 on launch day, but some routers with generic aliases like “grok-latest” may take a while to point to the new version.
Impact and Analysis
The most significant part of the launch isn’t that Grok 4.6 wins across the board (it doesn’t), but that it closes the gap that separated xAI from the rest of the frontier labs in one specific area: agents that sustain context and execute steps over minutes or hours, not seconds. The jump from 47.1% to 57.5% on APEX-Agents, against 56.7% for GPT-5.6 Sol Max, is the data point that matters most for anyone evaluating which model to put behind a production agent.
For the Spanish-language development ecosystem, this translates into a real alternative to other frontier models within tools like Cursor, without having to change workflows: the same editor, the same chat protocol, a different model behind it. Simultaneous availability on OpenRouter also reduces the friction of trying it from any existing stack without renegotiating contracts directly with xAI.
The honest limitation remains Terminal-Bench: a score of 26% leaves Grok 4.6 out of contention against rivals that top 34% on the same test. If your use case is, for example, a DevOps agent that executes infrastructure commands autonomously, it’s worth testing that specific scenario before migrating to production, regardless of what the general index average says.
What’s Next for Grok 4.6
xAI didn’t announce a date for a smaller or lower-cost version of Grok 4.6, something rivals have done with small variants of their frontier models. With prices of $2 and $6 per million tokens, Grok 4.6 lands in a range similar to GPT-5.6 Sol Max, suggesting the next pricing battle will play out in the fast variants and in partner volume discounts (OpenRouter, Vercel, Cloudflare) rather than in the base model.
The double-usage promotion in Cursor and Grok Build lasts only the first week, so the window to evaluate the model at no extra cost in real workflows is short. It’s worth taking advantage of it with a concrete use case and a real repository, not a demo, before it ends.
📖 Summary on Telegram: View summary
Try it yourself: run curl -fsSL https://x.ai/cli/install.sh | bash today and point your first code agent at grok-4.6 before the double-usage week ends in Cursor and Grok Build.
Frequently Asked Questions
Does Grok 4.6 replace Grok 4.5?
Yes, xAI presents it as the direct successor. Grok 4.5 remains available on some platforms, but the company’s active development now focuses on the 4.6 line.
How much does it cost to use Grok 4.6 via API?
$2 per million input tokens and $6 per million output tokens on the standard variant. The fast variant costs double per token.
Is Grok 4.6 better than GPT-5.6 Sol Max?
It depends on the benchmark. They tie on the AA Intelligence Index with 61 points each, and Grok 4.6 wins on APEX-Agents and AA-Briefcase, but GPT-5.6 Sol Max clearly beats it on DeepSWE v1.1 and Terminal-Bench v3.0.
Can I use Grok 4.6 without going through the xAI API?
Yes. It’s been available since launch on OpenRouter, Vercel, and Cloudflare, in addition to being natively integrated into Cursor and Grok Build.
What does it mean that the model does self-evaluation?
xAI reports that, on multi-step tasks, Grok 4.6 reviews its own work before continuing, a behavior that emerges from reinforcement learning training on long agentic tasks, not from an explicit user instruction.
Do I need to install anything to try it?
Not necessarily. If you already use Cursor, the model shows up directly in the selector. For command-line use, the Grok Build installer runs the same way on macOS, Linux, and inside WSL on Windows.
References
- Introducing Grok 4.6: xAI’s official announcement with benchmarks, pricing, and availability.
- xAI API Docs: official documentation for integrating Grok 4.6 via API.
- OpenRouter: one of the partners serving Grok 4.6 on launch day.
- xAI (Wikipedia): background on the company behind the Grok model family.
📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Ecliptic Graphic en Unsplash
0 Comments