Overview
On September 19, 2025, xAI quietly announced Grok 4 Fast on X: a frontier reasoning model built for extreme cost efficiency, described by the company as the product of large-scale reinforcement learning aimed at maximizing "intelligence density."
Key Specs and Performance
- Context window: 2 million tokens
- Pricing: $0.20/M input tokens, $0.50/M output tokens — roughly 47x cheaper than Grok 4
- Unified architecture: seamlessly switches between reasoning and non-reasoning modes
- GPQA Diamond: 85.7% (Grok 4: 87.5%), using ~40% fewer thinking tokens
- AIME 2025 (no tools): 92.0% (Grok 4: 91.7%, Grok 3 Mini: 83.0%)
- HMMT 2025: 93.3%
- LiveCodeBench: 80.0%, edging out Grok 4 (79.0%)
- Speed: 344 tokens/second output, ~3.8s end-to-end latency (pre-release API tests), about 2.5x faster than GPT-5's API
- Artificial Analysis: intelligence index score of 60, matching Gemini 2.5 Pro and Claude 4.1 Opus at ~1/25 the cost; completed the index evaluation with 61M tokens vs. Gemini 2.5 Pro's 93M
- xAI: Grok 4 Fast announcement
- Lech Mazur's NYT Connections Benchmark
- Artificial Analysis review
- Benchmark leaderboard post
- Multimodal demo
Access is available via grok.com, x.com, and Grok mobile apps (Grok 3 free with limits; Grok 4 Fast requires SuperGrok or PremiumPlus subscriptions). The model also integrates web and X search for real-time information, ranking 1163 Elo in the LMArena search arena.
Extended NYT Connections Benchmark
The standard NYT Connections benchmark (436 puzzles) had been saturated after o1 hit 90.7%. Lech Mazur extended it on GitHub to 759 puzzles, adding up to four carefully vetted "decoy" words per puzzle that don't belong to any category but lure models into wrong groupings. Models must form all four correct groups within four mistakes; the latest 100 puzzles are reserved to guard against training-data contamination.
Leaderboard highlights:
| Model | Score | |---|---| | Grok 4 Fast Reasoning | 92.1% | | Grok 4 | 91.7% | | GPT-5, o3-pro (medium), Gemini 2.5 Pro, DeepSeek, Qwen 3 | trailing | | Human average | ~71% (98% easy days, 39% hard) | | o1 | 98.9% (top AI on standard set) |
The decoy words increase task entropy, forcing genuine multi-step pattern recognition rather than intuitive guessing.
Multimodal Demo and Community Reactions
A widely shared X post showed Grok 4 Fast correctly solving a visual puzzle—inferring the fill order (7, 6, 3) of partially occluded glasses. Commenters questioned whether it had memorized training data; the poster clarified it is a fully retrained model and shouldn't be compared to older Grok versions. Others noted the milestone of frontier-level intelligence being commoditized: "18 months ago GPT-4-level reasoning was scarce and expensive; now Grok 4 Fast is fast, cheap, and efficient, opening real-time high-volume use cases like customer service and coding assistants."
Takeaways
1. Grok 4 Fast sets a new price-performance frontier, making frontier-level reasoning accessible for real-time, high-volume applications. 2. The extended Connections benchmark shows decoy-heavy word puzzles still differentiate top models, with xAI's models now leading. 3. Open questions remain around speed under load, non-reasoning-mode performance, and benchmark contamination—community scrutiny continues.