Model Drop

Grok 4 Fast

September 19, 2025 2 min read

Grok 4 Fast, which xAI released on September 19, 2025, is a cheaper, faster reasoning model built on Grok 4's training that runs reasoning and quick replies on the same weights with a 2M-token context window.

Model: Grok 4 Fast (grok-4-fast-reasoning, grok-4-fast-non-reasoning)
Maker: xAI
Ship date: September 19, 2025
Pricing: $0.20 / $0.50 per M tokens (prompts under 128K; $0.40 / $1.00 above)
Context: 2M tokens
Weights: Closed
Where you can use it:
- Grok and the Grok apps, free users included
- The xAI API, OpenRouter and Vercel AI Gateway
When to reach for it:
- Search-heavy questions on a budget
- Skip it for the hardest reasoning tasks
How it compares:
- Cheaper than GPT-5 mini and Gemini 2.5 Flash
What makes it different:
- One set of weights for both modes
- Uses 40% fewer thinking tokens than Grok 4
Early reactions: Early users praised the price-to-performance, with one calling it Gemini 2.5 Pro-level intelligence at about 25 times lower cost.
Read more: Grok 4 Fast and the Grok 4 Fast Model Card

The benchmarks

xAI reports pass@1 for Grok 4 Fast with reasoning against Grok 4, Grok 3 Mini (high), GPT-5 (high) and GPT-5 mini (high), with no tools on the math and Humanity's Last Exam rows; the search rows (BrowseComp, SimpleQA) let it browse the web and X and compare only against xAI's own models. The LMArena row uses LMArena's grok-4-fast-chat entry; the separate grok-4-fast-reasoning entry debuted at 1396 in the same snapshot.

xAI's numbers

GPQA Diamond (grad-level science): 85.7%. −1.8 under Grok 4 (87.5%), +0.0 over GPT-5 (high) (85.7%), +3.4 over GPT-5 mini (high) (82.3%), +6.7 over Grok 3 Mini (high) (79.0%).
AIME 2025 (math contest, no tools): 92.0%. +0.3 over Grok 4 (91.7%), −2.6 under GPT-5 (high) (94.6%), +0.9 over GPT-5 mini (high) (91.1%).
HMMT 2025 (math contest, no tools): 93.3%. +3.3 over Grok 4 (90.0%), +0.0 over GPT-5 (high) (93.3%), +5.5 over GPT-5 mini (high) (87.8%).
Humanity's Last Exam (no tools): 20.0%. −5.4 under Grok 4 (25.4%), −4.8 under GPT-5 (high) (24.8%), +3.3 over GPT-5 mini (high) (16.7%).
LiveCodeBench (coding problems, Jan to May): 80.0%. +1.0 over Grok 4 (79.0%), −6.8 under GPT-5 (high) (86.8%), +2.6 over GPT-5 mini (high) (77.4%).
BrowseComp (web research agent): 44.9%. +1.9 over Grok 4 (43.0%).
SimpleQA (short facts, with search): 95.0%. +1.0 over Grok 4 (94.0%), +13.0 over Grok 3 (82.0%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 17.9. +1.1 over GPT-5 mini (high) (16.8), +4.8 over Gemini 2.5 Flash (13.1).
LMArena text Elo (grok-4-fast-chat entry, debut): 1425. +18 over Gemini 2.5 Flash (1407), +32 over GPT-5 mini (high) (1393).
ARC-AGI-1 (ARC Prize, semi-private set): 48.5%. −5.83 under GPT-5 mini (high) (54.33%).
ARC-AGI-2 (ARC Prize, semi-private set): 5.28%. +0.84 over GPT-5 mini (high) (4.44%).