Xai launches GROK-4-FAST: unified inference and irrational models with 2M token context, and end-to-end trained end-to-end with tool-using reinforcement learning (RL)

Xai introduced it grok-4-fasta cost-optimized successor to GROK-4, who fuses “inference” and “irrational” behavior into a single set of weights that can be controlled via a system prompt. The model targets high-throughput search, coding, and Q&A 2m token context window Native Tool Use RL that determines when to browse the web, execute code, or invoke tools.

Architecture Notes

Previous GROK divides long chain “inference” and short “irrational” responses across separate models. Grok-4-Fast's Uniform heavy space Steering action reduces end-to-end latency and tokens via system prompts. This is related to real-time applications (search, assistant agents, interactive coding) where switching models penalize both latency and cost.

Search and use agents

Grok-4-Fast was trained end-to-end Tool use reinforcement learning Shows the benefits of search-centric agent benchmarks. BrowseComp44.9%, SimpleQA 95.0%, Recali Search 66.0%plus higher scores for Chinese variations (e.g.: browsecomp-zh51.2%). Xai also quotes private battle tests on Lmarena grok-4-fast-search (codename “Menro”) ranked #1 in Search Arena on 1163 ELOand a text variant (codename “Tahoe”). #8 in the field of textalmost the same grok-4-0709.

Performance and efficiency delta

Internal and public benchmarks, GROK-4-FAST posts Frontier class score While disconnecting the use of tokens. XAI report passes the result of @1 92.0% (AIME 2025, no tools), 93.3% (HMMT 2025, no tools), 85.7% (GPQA diamond)and 80.0% (livecodebench January to May)it approaches or matches the Grok-4, but is using it About 40% less “thinking” tokens On average. The company claims this is “intelligence density.” ~98% reduction in price to reach the same benchmark performance as GROK-4 When lower token counts combined with pricing per new token.

Development and price

The model is Usually available to all users At Grok's fast and Automatic Web and mobile modes. Auto selects GROK-4-FAST for difficult queries to improve latency without losing quality.Free users Access Xai's latest model tier. For developers, Xai will publish it Two SKUs–grok-4-fast-reasoning and grok-4-fast-non-reasoning– Which 2M Context. Price (Xai API) is $0.20/1m input token (<128k), $0.40/1m input token (≥128k), $0.50/1m ​​output token (<128k), $1.00/1m output token (≥128k)and $0.05/1M cache input token.

5 Technical Takeout:

• Unified model + 2M context. Grok-4-Fast uses a single heavy space for “inference” and “irrational” with windows of 2,000,000 tokens in both Skus.
• Scale pricing. API pricing begins $0.20/m input, $0.50/m outputthere is cached input 0.05 dollars/m Higher rates above 128K context.
• Efficiency billing. Xai reports40% less “thinking” tokens Equivalent accuracy and GROK-4, ~98% lower price according to GROK-4 performance Frontier benchmark.
• Benchmark profile. Reported path @1: AIME-2025 92.0%, HMMT-2025 93.3%, GPQA-Diamond 85.7%, livecodebench (January – May) 80.0%.
• Using Agent/Search. Tool Use After Training Using RL. It is deployed for browsing/search workflows with documented search agent metrics and live search requests on documents.

summary

GROK-4-FAST Package GROK-4 level features become a single prompt stearable model with 2M token windows, tool-using RL, and pricing tailored for high-throughput search and agent workloads. The early public signals (lmarena #1 during search, competitive text placement) coincides with the similar accuracy claims of xai, translated to 40% less “thinking” tokens and reduced unit costs in production.

---

Please check Technical details. Please feel free to check GitHub pages for tutorials, code and notebooks. Also, please feel free to follow us Twitter And don't forget to join us 100k+ ml subreddit And subscribe Our Newsletter.