xAI
xAI releases Grok 4.6 for long-running agents
$2/$6 per million tokens, index parity with GPT-5.6 Sol, terminal use still weak.
The answer
xAI released Grok 4.6 on 12 August 2026 at $2/$6 per million tokens.
xAI released Grok 4.6 on 12 August 2026. The launch positions it for long-running agents and ambitious interactive and visual work rather than as a raw capability increase over Grok 4.5.
builds on Grok 4.5 with a particular focus on long-running agents and more ambitious interactive and visual work
Pricing and benchmarks
Pricing starts at $2 per million input tokens and $6 per million output, with a fast variant at twice the price. Below 200K prompt tokens, cached input is $0.50 per million. It is available in the API and via OpenRouter, Vercel and Cloudflare.
| Model | AA Index | Input / output |
|---|---|---|
| Claude Opus 5 | 63 | $5 / $25 |
| GPT-5.6 Sol | 61 | $5 / $30 |
| Grok 4.6 | 61 | $2 / $6 |
Grok 4.6 posted a GDPval-AA v2 Elo of 1753 and 15.8% on Harvey's Legal Agent Benchmark at launch. Terminal-Bench 3.0 was 26%. On xAI's own ten-row comparison, Claude Fable 5 wins the most rows; Grok is strongest on knowledge work and legal reasoning.
Efficiency and successor
On long agentic tasks the model completed work in roughly 53 turns at $0.84 per task, against Claude Opus 5's 103 turns. Artificial Analysis notes it starts answering later and can cost more per measured task than Grok 4.5 in some conditions.
Nvidia stated Grok 4.6 was trained and runs on GB300 NVL72 with NVLink. xAI did not publish a parameter count in the launch post or API documentation.
Grok 4.7, described by Musk as a 2.1-trillion-parameter successor, remained unreleased as of mid-September after delays attributed to reinforcement-learning tuning, response-length penalties and self-checking. On 13 September Musk named Grok 4.8 as a 2.5-trillion-parameter model on a new C++ stack.
On 24 July Musk said 4.7 would arrive in four weeks; four days later he revised that to a few weeks. On 11 September he said the model needed a few more days of tuning, noting it stopped too early on difficult tasks and did not check its own work rigorously enough. On 14 September he graded it as roughly on par with Opus 5.0 rather than 5.1.
All Grok 4.7 details trace to Musk's posts on X rather than xAI documentation. No context window, pricing, benchmark package or API model ID has been published. Grok 4.6 remains the latest officially available model in the line.
SpaceX, which bought xAI in February 2026, disclosed in its Nasdaq listing filing that the xAI segment lost $2.47bn from operations in the first quarter of 2026 on $818m of revenue. The same filing shows Anthropic paying SpaceX $1.25bn a month for Colossus compute — a rival lab renting the machines.
Frequently asked questions
What does Grok 4.6 cost?
How does it compare on benchmarks?
Is Grok 4.7 out?
Sources
- Introducing Grok 4.6 — xAI, 12 August 2026
- Grok 4.6: Features, Benchmarks, Pricing & Comparisons — DataCamp, 13 August 2026
- xAI's Grok 4.6: Benchmarks, Pricing, and the Full Developer Ecosystem — Better Stack, 14 August 2026
- xAI delays Grok 4.7 as Musk points to reinforcement-learning and self-checking problems — DataStudios, 14 September 2026
- Space Exploration Technologies Corp — Form S-1 (FY2026) — U.S. Securities and Exchange Commission, 20 May 2026
- Anthropic will pay xAI $1.25B per month for compute — TechCrunch, 20 May 2026