Model Drop

Grok 4.6

August 12, 2026 2 min read

Grok 4.6, which SpaceXAI released on August 12, 2026, is a follow-up to Grok 4.5 tuned for long-running agents and for visual, interactive work.

Model: Grok 4.6 (grok-4.6)
Maker: SpaceXAI
Ship date: August 12, 2026
Pricing: $2 / $6 per M tokens
Context: 500K tokens
Where you can use it:
- Cursor and Grok Build
- The xAI API, OpenRouter, Vercel, Cloudflare
When to reach for it:
- Long agent runs and office work on a budget
- Skip it for the hardest repo-level coding
How it compares:
- Level with GPT-5.6 Sol on office work
- Behind Claude Opus 5 on most coding tests
What makes it different:
- Grok 4.5 regenerated its training data
- Extra training on Cursor workflow data
Early reactions: Hacker News called Grok real competition at last but doubted the benchmarks and the one-pass design claims.
Read more: Introducing Grok 4.6 and the Grok 4.6 Model Card

The benchmarks

SpaceXAI ran Grok 4.6 at high effort against Grok 4.5 at high and GPT-5.6 Sol and Claude Fable 5 at max, using the best self-reported or public rival scores; Claude Opus 5, which leads most of these rows, appears only in the model card, not the launch post. Several benchmarks are run by SpaceXAI's partners (Cursor, Artificial Analysis, Mercor, Vals AI). The independent rows also use Grok 4.6 at high and rivals at max.

SpaceXAI's numbers

GDPval-AA v2 (professional work products): 1753 Elo. +227 over Grok 4.5 (1526), +25 over GPT-5.6 Sol (1728), +12 over Claude Fable 5 (1741), −96 under Claude Opus 5 (1849).
AA-Briefcase (multi-week office projects): 1577 Elo. +264 over Grok 4.5 (1313), +75 over GPT-5.6 Sol (1502), +3 over Claude Fable 5 (1574), −138 under Claude Opus 5 (1715).
APEX-Agents (cross-app professional tasks): 57.5%. +10.4 over Grok 4.5 (47.1%), +0.8 over GPT-5.6 Sol (56.7%), −1.7 under Claude Fable 5 (59.2%), −3.1 under Claude Opus 5 (60.6%).
Harvey Legal Agent Benchmark (legal work): 15.8%. +2.9 over Grok 4.5 (12.9%), +13.3 over GPT-5.6 Sol (2.5%), +4.5 over Claude Fable 5 (11.3%), +4.1 over Claude Opus 5 (11.7%).
CursorBench 3.2 (coding tasks in Cursor): 69.9%. +3.2 over Grok 4.5 (66.7%), +2.7 over GPT-5.6 Sol (67.2%), −0.6 under Claude Fable 5 (70.5%).
DeepSWE v1.1 (long software tasks): 65.9%. +11.9 over Grok 4.5 (54%), −7.1 under GPT-5.6 Sol (73%), −4.1 under Claude Fable 5 (70%), −8.1 under Claude Opus 5 (74.0%).
Terminal-Bench 3.0 (hard terminal tasks): 26%. +10.3 over Grok 4.5 (15.7%), −8.6 under GPT-5.6 Sol (34.6%), −8.1 under Claude Fable 5 (34.1%), −17.5 under Claude Opus 5 (43.5%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 44.3. +5.5 over Grok 4.5 (38.8), −2.7 under GPT-5.6 Sol (47.0), −5.3 under Claude Fable 5 (49.6), −6.5 under Claude Opus 5 (50.8).
ARC-AGI-2 (ARC Prize, semi-private set): 65.1%. +12.5 over Grok 4.5 (52.6%), −27.4 under GPT-5.6 Sol (92.5%), −24.1 under Claude Fable 5 (89.2%), −25.3 under Claude Opus 5 (90.4%).