{"id":39490,"date":"2025-09-21T06:40:26","date_gmt":"2025-09-21T06:40:26","guid":{"rendered":"https:\/\/youzum.net\/xai-launches-grok-4-fast-unified-reasoning-and-non-reasoning-model-with-2m-token-context-and-trained-end-to-end-with-tool-use-reinforcement-learning-rl\/"},"modified":"2025-09-21T06:40:26","modified_gmt":"2025-09-21T06:40:26","slug":"xai-launches-grok-4-fast-unified-reasoning-and-non-reasoning-model-with-2m-token-context-and-trained-end-to-end-with-tool-use-reinforcement-learning-rl","status":"publish","type":"post","link":"https:\/\/youzum.net\/fr\/xai-launches-grok-4-fast-unified-reasoning-and-non-reasoning-model-with-2m-token-context-and-trained-end-to-end-with-tool-use-reinforcement-learning-rl\/","title":{"rendered":"xAI launches Grok-4-Fast: Unified Reasoning and Non-Reasoning Model with 2M-Token Context and Trained End-to-End with Tool-Use Reinforcement Learning (RL)"},"content":{"rendered":"

xAI introduced Grok-4-Fast<\/a><\/strong>, a cost-optimized successor to Grok-4 that merges \u201creasoning\u201d and \u201cnon-reasoning\u201d behaviors into a single set of weights controllable via system prompts. The model targets high-throughput search, coding, and Q&A with a 2M-token context window<\/strong> and native tool-use RL that decides when to browse the web, execute code, or call tools.<\/p>\n

Architecture note<\/strong><\/h3>\n

Previous Grok releases split long-chain \u201creasoning\u201d and short \u201cnon-reasoning\u201d responses across separate models. Grok-4-Fast\u2019s unified weight space<\/strong> reduces end-to-end latency and tokens by steering behavior via system prompts, which is relevant for real-time applications (search, assistive agents, and interactive coding) where switching models penalizes both latency and cost.<\/p>\n

Search and agentic use<\/strong><\/h3>\n

Grok-4-Fast was trained end-to-end with tool-use reinforcement learning<\/strong> and shows gains on search-centric agent benchmarks: BrowseComp 44.9%<\/strong>, SimpleQA 95.0%<\/strong>, Reka Research 66.0%<\/strong>, plus higher scores on Chinese variants (e.g., BrowseComp-zh 51.2%<\/strong>). xAI also cites private battle-testing on LMArena where grok-4-fast-search<\/code> (codename \u201cmenlo\u201d) ranks #1 in the Search Arena with 1163 Elo<\/strong>, and the text variant (codename \u201ctahoe\u201d) sits at #8 in the Text Arena<\/strong>, roughly on par with grok-4-0709<\/code>. <\/p>\n

Performance and efficiency deltas<\/strong><\/h3>\n

On internal and public benchmarks, Grok-4-Fast posts frontier-class scores<\/strong> while cutting token usage. xAI reports pass@1 results of 92.0% (AIME 2025, no tools)<\/strong>, 93.3% (HMMT 2025, no tools)<\/strong>, 85.7% (GPQA Diamond)<\/strong>, and 80.0% (LiveCodeBench Jan\u2013May)<\/strong>, approaching or matching Grok-4 but using ~40% fewer \u201cthinking\u201d tokens<\/strong> on average. The company frames this as \u201cintelligence density,\u201d claiming a ~98% reduction in price to reach the same benchmark performance as Grok-4<\/strong> when the lower token count and new per-token pricing are combined.<\/p>\n

Deployment and price<\/strong><\/h3>\n

The model is generally available to all users<\/strong> in Grok\u2019s Fast<\/strong> and Auto<\/strong> modes across web and mobile; Auto will select Grok-4-Fast for difficult queries to improve latency without losing quality, and\u2014for the first time\u2014free users<\/strong> access xAI\u2019s latest model tier. For developers, xAI exposes two SKUs<\/strong>\u2014grok-4-fast-reasoning<\/code> and grok-4-fast-non-reasoning<\/code>\u2014both with 2M context<\/strong>. Pricing (xAI API) is $0.20 \/ 1M input tokens (<128k)<\/strong>, $0.40 \/ 1M input tokens (\u2265128k)<\/strong>, $0.50 \/ 1M output tokens (<128k)<\/strong>, $1.00 \/ 1M output tokens (\u2265128k)<\/strong>, and $0.05 \/ 1M cached input tokens<\/strong>.<\/p>\n

\n
\"\"
https:\/\/x.ai\/news\/grok-4-fast<\/figcaption><\/figure>\n<\/div>\n

5 Technical Takeaways:<\/strong><\/h3>\n