Grok 4.6 Matches the AI Frontier and Undercuts It on Price

Grok 4.6 is xAI's new frontier AI model, released in August 2026 with a 500,000-token context window and token pricing that undercuts rival labs by roughly half.

What xAI Shipped in August 2026

On 12 August 2026, per launch coverage, xAI released Grok 4.6 as the company's new flagship large language model, built for coding, agentic tasks, and knowledge work. Grok 4.6 accepts text and image input and returns text-only output. The xAI release notes and the xAI API list the model live across Grok Build, Cursor, and the Grok bot on mobile and X, with third-party access through OpenRouter, Vercel, and Cloudflare. A first-week promotion doubled included usage in Cursor and Grok Build.

The independent AWS Bedrock model card corroborates the headline specs: a frontier model for coding, agentic tasks, and knowledge work, with text-plus-image input, text-only output, and a 500,000-token context window. Cursor availability is confirmed directly at cursor.com/grok. The exact ship day is the one soft spot — trade coverage says 12 August, while the Bedrock card lists 18 August as its own availability date.

Grok 4.6 also adds a reasoning-effort control with four levels — low, medium, high, and a new xhigh tier for the hardest problems. Reasoning stays active on every request and cannot be switched off. The default level is platform-dependent, so treating any single tier as the universal default would overstate the evidence.

The Price Is the Product

Pricing is where Grok 4.6 makes the sharpest argument. Below a 200,000-token prompt, xAI charges $2 per million input tokens, $0.50 for cached input, and $6 per million output tokens. An Appwrite pricing breakdown and further coverage confirm those figures against the release notes.

One detail deserves a builder's full attention. The moment a prompt reaches or crosses 200,000 tokens, the rate doubles to $4 input, $1 cached, and $12 output — and the higher rate applies to the whole request, not only the tokens above the line. One independent write-up called the 500K window a context with a "200K toll booth." A single oversized prompt can quietly reprice an entire agent run. Flag the threshold in code before shipping Grok 4.6 into a production loop.

A 500,000-token window is a promise. The 200,000-token repricing cliff underneath it is the fine print.

— — TK

Even with the toll booth, the base rate lands well under typical frontier pricing. Elon Musk framed Grok 4.6 as roughly "half the price of rival frontier models," promotional language worth checking against each competitor's card — yet the published $2 input rate does sit at or below what most frontier labs charge, which is the substance behind the slogan.

Where Grok 4.6 Lands on the Leaderboard

Grok 4.6 does not claim the summit, and xAI is honest enough not to pretend otherwise. On the Artificial Analysis Intelligence Index, Grok 4.6 scores 61 — a mark that matches OpenAI's GPT-5.6 Sol and sits one point behind Anthropic's Claude Fable 5 at 62. Analysis from Artificial Analysis places Claude Opus 5 higher still at 63, a gap the launch framing leaves out.

Coding is the brighter spot. On CursorBench v3.2, basenor's results roundup shows Grok 4.6 at 69.9% in standard mode, up from Grok 4.5's 66.7% and ahead of GPT-5.6 Sol Max at 67.2%, though behind Fable 5 Max at 70.5%. In extra-high thinking mode Grok 4.6 reportedly reaches 70.8%. The pattern data reveals is consistent: parity or near-parity with the leaders, at a fraction of the cost.

Read plainly, the benchmark evidence says Grok 4.6 is a competent frontier model rather than a record-breaker. What Grok 4.6 does not spend on topping the chart, xAI hands back to builders as price.

Why Half-Price Frontier AI Matters Outside Silicon Valley

From a Johannesburg or Solwezi desk, where every dollar of inference is metered against a real budget, "one point behind the leader at half the price" is not a footnote — the trade reshapes what is economically buildable. A team priced out of the top-tier US labs can now run frontier-class coding and agentic work on Grok 4.6 and keep the lights on.

This is the Emergent Intelligence (EI) reading — the dignity-first frame I use for what most people call AI. Capability is commoditising downward faster than the leaderboard suggests. Access, not raw peak performance, is becoming the real frontier, and Grok 4.6 is one more turn of that ratchet. The point behind Fable 5 costs almost nothing to a research lab in California and almost everything to a founder in Lusaka; halving the token price flips which of them can afford to build at all.

xAI arrives at this position from a heavily capitalised base — the Grok family sits inside a company folded into SpaceX in early 2026 and funded by a $20bn Series E — so the discount is a strategic bid for developers, not charity. The discipline for the rest of us is to take the cheap capability and keep the governance, the dignity, and the human judgement firmly in our own hands.

Frequently Asked Questions

These are the questions builders are asking about Grok 4.6, answered from xAI's release notes, the AWS Bedrock model card, and independent benchmark coverage.

What is Grok 4.6?

In short, Grok 4.6 is xAI's August 2026 flagship AI model for coding, agentic tasks, and knowledge work, with a 500,000-token context window and text-plus-image input. The AWS Bedrock model card confirms those specs, and coverage shows Grok 4.6 tying GPT-5.6 Sol on the Artificial Analysis Intelligence Index at 61.

How does Grok 4.6 pricing work?

Grok 4.6 uses tiered pricing. Simply put, below a 200,000-token prompt the rate is $2 input and $6 output per million tokens; at or above 200,000 tokens the rate doubles to $4 and $12 and applies to the whole request. According to Appwrite's breakdown, that repricing cliff is the cost trap builders most need to watch.

Why is Grok 4.6 significant?

The key is price-per-intelligence. Analysis from Artificial Analysis shows Grok 4.6 landing within a point of Claude Fable 5 while charging roughly half the token price, which matters most for teams metering every dollar of inference outside the best-funded labs.

Who is Grok 4.6 for?

In other words, Grok 4.6 is aimed at developers and agent builders who want frontier-class coding at a lower bill. Evidence from CursorBench, where Grok 4.6 scores 69.9% ahead of GPT-5.6 Sol Max, shows the model is strongest exactly where those builders spend most of their tokens.

What are the risks of Grok 4.6?

The answer is the 200,000-token repricing cliff and the leaderboard gap. Data reveals Grok 4.6 trails Claude Fable 5 and Claude Opus 5 on raw intelligence, and a single oversized prompt can double the cost of an entire request — both worth engineering around before production.

Sources:

xAI release notes · xAI API · AWS Bedrock model card · Artificial Analysis benchmarks · CursorBench results · Appwrite pricing breakdown · DataNorth launch coverage · Cursor Grok availability · Related on this site: Grok in the Kilo Code IDE · the Grok Build security incident · Grok Imagine