Grok 4.6: xAI's Agent-Focused Update Matches GPT-5.6 Sol at the Same $2/$6 Price

xAI released Grok 4.6 on August 12, 2026. The update builds on Grok 4.5 with a stated focus on long-running agents and "more ambitious interactive and visual work," and it lands at the exact same price: $2 per million input tokens, $6 per million output tokens. Per xAI, it matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index at 61, one point behind Fable 5 Max's 62.

This is the second xAI release in five days, after Grok Imagine Image 2.0, and the first model release from the family since Grok 4.5 launched in Cursor on July 9. The interesting signal is not the headline parity, it is where the benchmark deltas sit: agentic coding surfaces improved sharply, while the terminal-agent score still trails the frontier.

Official Sources#

What Shipped#

Grok 4.6 is a point release over Grok 4.5, not a new architecture generation. The model id on the API is grok-4.6, with a 500,000-token context window, text and image input, and text output with no output limit. Reasoning effort is selectable at low, medium, high (default), or xhigh.

Training changes per the announcement: a longer supplemental run than Grok 4.5, curated model-generated data for reasoning and technical concepts, and agentic RL on domain environments that include kernel optimization, web development, and computer-aided design. xAI says it used Grok 4.5 to regenerate SFT trajectories across reasoning efforts and agent harnesses, filtering problematic traces with model-based checks. On the safety side, xAI says Grok 4.6 shipped with its "widest-ever suite" of pre-deployment testing and that safeguards were calibrated to the expanded capabilities.

The Benchmarks#

xAI's published table compares Grok 4.6 High against Grok 4.5 High, GPT-5.6 Sol Max, and Fable 5 Max. Competitor figures are the vendors' own reported or leaderboard numbers, with that caveat stated by xAI itself:

Two readings matter. On CursorBench 3.2, Grok 4.6 (69.9%) is the highest of the four, beating GPT-5.6 Sol by 2.7 points and landing just behind Fable 5 Max's 70.5%. Grok 4.5's whole origin story was Cursor training, and 4.6 extends that edge. GDPVal-AA v2 (1753) and APEX-Agents (57.5%) also beat GPT-5.6 Sol. Given the model is priced at roughly a third of Fable 5's $10/$50, parity plus on the coding benchmarks is the value story.

The second reading is the lag: Terminal-Bench v3.0 at 26% is 8.6 points behind GPT-5.6 Sol and 8.1 behind Fable 5 Max, and DeepSWE v1.1 at 65.9% trails Sol by 7.1 points. For a release pitched at long-running agents, the gap on the terminal-work benchmark is the one number to keep an eye on in independent evals, because it is exactly the workload our own Terminal-Bench analysis found predicts agent quality best.

Pricing#

The price card is unchanged from Grok 4.5, with a long-context multiplier for requests over 200K tokens:

There is also a fast variant at twice the price, per the announcement. At $2/$6, Grok 4.6 sits in the same tier as GPT-5.6 Luna and DeepSeek V4 Pro on input, and above DeepSeek V4 Flash's $0.14/$0.28 but far below Fable 5's $10/$50 - the spread we broke down in Fable 5 vs DeepSeek V4: cost vs quality. xAI recommends setting a prompt_cache_key for agent loops so cache hits are reliable across requests; the cache-read rate is $0.50 per 1M, a 75% input discount that matters for long multi-turn sessions.

Run It in OpenCode#

Grok 4.6 is available in OpenCode today as opencode/grok-4.6, registered at the same $2/$6/$0.50 pricing as the first-party API. Install with the official one-liner from the OpenCode docs:

curl -fsSL https://opencode.ai/install | bash

# One-shot run at default (high) reasoning
opencode run --model opencode/grok-4.6 \
  "analyze this repo, find the memory leak, and open a fix"

# Interactive session with the model preselected
opencode --model opencode/grok-4.6

Outside OpenCode, the model is available in Cursor and Grok Build on day one (xAI is offering 2x included usage in both for the first week), on the xAI API at console.x.ai, and through OpenRouter, Vercel, and Cloudflare. On Vercel's AI Gateway the model id is xai/grok-4.6, with the same 500K context, image input, and low/medium/high/xhigh reasoning levels, callable through the AI SDK.

Grok 4.6 vs the Field#

Pick Grok 4.6 when: the workload is agentic coding with strong first-pass results - CursorBench is the headline strength; the price is right for fleets; you want 500K context with image input for repo screenshots or UI work.

Wait or pick GPT-5.6 Sol / Fable 5 when: the workload leans terminal-heavy (long shell loops, multi-command debugging), where the Terminal-Bench and DeepSWE deltas are largest, or when the task is frontier reasoning where Fable 5's 62 on the AA Intelligence Index is the top of the table.

For how Grok 4.6 stacks against the open-weight tier, the GLM 5.2 vs DeepSeek V4 vs Qwen3 showdown remains the reference, and our Grok 4.5 developer guide covers the setup details that carry over unchanged.

FAQ#

What is Grok 4.6?#

The August 12, 2026 point release from xAI, focused on long-running agents and interactive/visual work. It matches GPT-5.6 Sol on the AA Intelligence Index at 61 and keeps Grok 4.5's $2/$6 per million token pricing. Model id on the API: grok-4.6.

How much does Grok 4.6 cost?#

$2.00 per million input tokens and $6.00 per million output tokens for short context (under 200K tokens). Long-context requests (200K-500K) cost $4.00 input and $12.00 output. Cached input is $0.50 (short) or $1.00 (long) per million tokens. A fast variant is priced at 2x.

What is the context window?#

500,000 tokens, with text and image input and text output. There is no output token limit.

Where can I run Grok 4.6?#

Cursor and Grok Build (2x included usage for the first week), the xAI API, OpenRouter, Vercel AI Gateway, and Cloudflare. In OpenCode it is available now as opencode/grok-4.6.

How does Grok 4.6 compare to Grok 4.5?#

It is the same architecture family with a longer supplemental training run and agentic RL on kernel, web, and CAD environments. Against 4.5 High it is up on every published benchmark, most sharply on CursorBench 3.2 (69.9% vs 66.7%) and Terminal-Bench v3.0 (26% vs 15.7%).

Sources#

Last updated: August 13, 2026

Continue Reading#

• Grok 4.5 for Developers - the previous release's setup guide, still valid for 4.6
• Grok 4.5 in 10 Minutes - video companion covering the 500K context and Build-mode workflow
• DeepSeek V4 Flash 0731 OpenCode Guide - the OpenCode-centered release format for the value tier
• GLM 5.2 in 9 Minutes - Zhipu's open-weight alternative at a fraction of the price
• Fable 5 vs DeepSeek V4: Cost vs Quality - when the cheap model is the right model
• Grok Code Fast 1: xAI''s Speed-Optimized Coding Model