xAI

Grok 4.6

Released Aug 12, 2026 · knowledge cutoff 2026-02

Status: Superseded
Location: United States
Type: LLM
Modality: Multimodal
Context window: 500K
Max output: —
Speed (high effort): 65 tok/s · 32s to first answer token
Price ($/MTok in / out): $2 / $6
Cost per task (high effort): $0.84
Open weights: No

Benchmarks3 of 10 reported

Terminal-Bench 2.1

88.4%

GDPval-AA v2

1753

ARC-AGI-2

67.1%

MMLU-Proretired

—

GPQA Diamond

—

SWE-bench Verified

—

SWE-bench Pro

—

AIME (latest)retired

—

Humanity's Last Exam

—

LMArena Elo

—

Strengths & weaknesses

Strengths

Sustains a long agentic run without losing the thread — the work went into fine-tuning on regenerated trajectories rather than a new base model
Reasons through dense documents rather than pattern-matching them, which shows in legal and knowledge work
Holds the previous generation's price despite the capability step, so the cost of a finished task fell outright

Weaknesses

Very slow to the first answer token — the thinking phase runs for half a minute before any output appears
Weakest on terminal and command-line work, well behind its scores on other agentic tasks
Crossing the 200K-token prompt threshold reprices the whole request at double rate, so long-context runs cost more than the headline suggests

Reuses Grok 4.5's 1.5T-parameter base, with the gains from supervised fine-tuning on regenerated trajectories and reinforcement learning rather than a new pre-train. Terminal-Bench 2.1 88.4% is Artificial Analysis-measured; xAI's own card leads with Terminal-Bench 3.0 (26.5%), a different and much harder version that is not comparable to the 2.1 figure this site tracks. GPQA Diamond, MMLU-Pro, HLE, SWE-bench Verified and AIME are unpublished by both xAI and AA (ARC-AGI-2 was later verified by ARC Prize, below), and the model has no arena.ai listing, so lmarenaElo is null — note that the widely-quoted "1753 Elo" is GDPval-AA v2, not LMArena. AA Intelligence Index 61 (joint third at release, level with GPT-5.6 Sol), cost per task $0.84 and speed measured at high effort. Pricing $2/$6 applies below a 200K-token prompt; above it every token in the request bills at $4/$12, and a fast variant costs double. xAI publishes no max-output figure. The company rebranded from xAI to SpaceXAI on 2026-07-06 after merging into SpaceX; companies.json still carries the old name. ARC-AGI-2 67.1% is ARC Prize-verified at xhigh effort ($0.76/task; high 65.1%, medium 61.3%), published 2026-08-11 at arcprize.org/results/xai-grok-4-6.

For developers

API model strings

xaigrok-4.6

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor