Grok 4 Fast: 2M Context, Pricing & Deprecation (2026)
Grok 4 Fast is xAI's 2 million token context model launched Sept 19, 2025, scoring 35 on the AA Intelligence Index. Now deprecated, retiring Aug 15, 2026.
Grok 4 Fast is xAI's cost-optimized 2 million token context model, released September 19, 2025, built for large-document analysis and agentic tool-calling. Its grok-4-fast-search variant ranked first in the LMArena Search Arena at 1163 Elo, but xAI deprecated the whole model on May 15, 2026 ahead of an August 15, 2026 retirement.
Grok 4 Fast, released by xAI on September 19, 2025, is a 2 million token context model scoring 35 on the Artificial Analysis Intelligence Index. It uses 40% fewer reasoning tokens than Grok 4 for comparable benchmark results, but was deprecated May 15, 2026 in favor of Grok 4.3.
Provider: xAI · Family: Grok 4
Context window: 2,000,000 tokens · Max output: 16,000
Input modalities: text, image, tool-calls · Output: text, tool-calls
About Grok 4 Fast
Grok 4 Fast is a large language model built by xAI, the Elon Musk-founded AI lab now merged into SpaceX as SpaceX-xAI. It launched on September 19, 2025 as a speed- and cost-optimized sibling to the flagship Grok 4, which shipped two months earlier on July 9, 2025. Architecturally it is a Mixture-of-Experts Transformer derived from the Grok 4 base, distilled to use roughly 40 percent fewer reasoning tokens per response while targeting comparable benchmark performance. xAI ships it as two separate API model strings rather than one model with a toggle: grok-4-fast-reasoning for extended chain-of-thought tasks and grok-4-fast-non-reasoning for low-latency chat and retrieval. On independently measured benchmarks, Grok 4 Fast Reasoning scores 35 on the Artificial Analysis Intelligence Index, a composite covering reasoning, knowledge, math, and coding, versus 33 for the base Grok 4 and 39 for the later Grok 4.1 Fast. xAI does not separately publish per-model GPQA Diamond or AIME 2025 scores for the Fast variant, only the claim that it matches Grok 4's frontier benchmark results at roughly 2 percent of the cost. Its most distinctive independently verified result came in a narrower slice: the grok-4-fast-search variant ranked first in the LMArena Search Arena at 1163 Elo, 17 points ahead of OpenAI's o3-search, according to Artificial Analysis reporting. The headline spec is context: a 2,000,000 token window, among the largest of any production model at launch and still larger than GPT-4.1's 1M and Claude Sonnet 4.6's 200K. Despite that, the xAI playground caps a single response at 16,000 output tokens, so long-context use cases are about ingesting large inputs, not generating proportionally long outputs. xAI applies a higher, unpublished extended-context rate once a single request crosses the 128,000 token mark, a tiered-pricing pattern it shares with the base Grok 4 model. Modalities are text and image input with text output, plus native tool use. Through xAI's Agent Tools API, Grok 4 Fast can call live X search, general web search, and remote code execution mid-conversation, which is the main reason it gets used in agentic research and customer-support pipelines rather than as a plain chat model. It does not accept audio or video input and has no native audio or video output. Standard API pricing is $0.20 per 1 million input tokens and $0.50 per 1 million output tokens, with cached input tokens at $0.05 per 1 million, a 75 percent discount versus fresh input. A 500,000-token codebase summarization with a 20,000-token output costs roughly $0.11. A daily agentic coding loop pushing 2 million input tokens and 200,000 output tokens runs about $0.50. A support chatbot handling 1,000 turns at 2,000 input and 500 output tokens per turn costs around $0.65. These rates held steady through July 2026 even after the model's deprecation notice, unlike some vendors that discount models on the way out. Access is through the direct xAI API (api.x.ai, OpenAI-compatible SDK surface) and OpenRouter (x-ai/grok-4-fast). Oracle Cloud briefly listed it in its Generative AI catalog but now marks that listing deprecated too. There is no confirmed availability on AWS Bedrock or Google Vertex AI. Time to first token measured 0.64 seconds on the non-reasoning variant per Artificial Analysis, with throughput around 227 tokens per second in reasoning mode, and the API supports up to 4 million tokens per minute of throughput. xAI's published model card for Grok 4 Fast (dated September 19, 2025) describes standard safety training aimed at refusing requests that would lower the barrier to chemical, biological, radiological, or nuclear weapon development, along with refusals for CSAM and self-harm content. A system prompt layer adds further mitigations documented in that card. xAI also evaluates robustness to jailbreaks and prompt injection and reports working to reduce deceptive or power-seeking behavior, consistent with its safety framing for the rest of the Grok 4 family. Grok 4 Fast is the wrong choice for any new long-term integration: xAI deprecated it on May 15, 2026, and has scheduled a full API retirement for August 15, 2026, redirecting traffic toward Grok 4.3. It remains a reasonable fit only for short-term cost-sensitive workloads already built around it, agentic search and tool-calling tasks that lean on the Agent Tools API, and very large single-request document or codebase analysis where the 2M context window matters more than having the newest benchmark scores. Teams starting fresh should evaluate Grok 4.3 or Grok 4.5 instead, both of which supersede it on xAI's own migration guidance.
Pricing
$0.20 per 1M input tokens, $0.50 per 1M output tokens, and $0.05 per 1M cached input tokens via the xAI API since the September 19, 2025 launch. xAI applies a higher, unpublished extended-context rate once a single request crosses 128,000 tokens. Pricing was unchanged as of July 2026 despite the model's May 15, 2026 deprecation notice.
Key Features
- 2 Million Token Context Window: Holds roughly 1.5 million words of text, image, or code in a single request, larger than GPT-4.1's 1M and Claude Sonnet 4.6's 200K windows.
- Dual Reasoning Modes: Call grok-4-fast-reasoning for step-by-step chain-of-thought or grok-4-fast-non-reasoning for low-latency responses, using 40 percent fewer thinking tokens than Grok 4 on average.
- Agent Tools API Integration: Connects to live X search, general web search, and remote code execution so agents can pull real-time data mid-conversation.
- Sub-Second Time to First Token: Measured 0.64-second time to first token on the non-reasoning variant, per Artificial Analysis benchmarking.
- Prompt Caching: Cached input tokens cost $0.05 per 1M, a 75 percent discount off the standard $0.20 input rate, for repeat-context workloads.
Pros
- 2 million token context window is among the largest of any production model, holding an entire large codebase or long report in one request.
- $0.20/$0.50 per 1M input/output tokens delivers benchmark performance close to Grok 4 at roughly 2 percent of the cost, per xAI's own launch comparison.
- grok-4-fast-search topped the LMArena Search Arena leaderboard at 1163 Elo, 17 points ahead of OpenAI's o3-search, per Artificial Analysis.
Cons
- Deprecated by xAI on May 15, 2026, with full API retirement scheduled for August 15, 2026, so new projects should target Grok 4.3 instead.
- Independent per-benchmark scores such as GPQA and AIME specific to this model are not separately published, only a composite Artificial Analysis Intelligence Index of 35.
- Playground responses are capped at 16,000 output tokens despite the 2 million token context window.
Benchmarks
- artificial analysis intelligence index: 35
- artificial analysis speed tokens per sec: 227
Frequently Asked Questions
What is Grok 4 Fast and who built it?
Grok 4 Fast is a large language model built by xAI, the AI lab founded by Elon Musk and now merged into SpaceX. It launched on September 19, 2025 as a faster, cheaper sibling to the flagship Grok 4, which shipped two months earlier. Architecturally it is a Mixture-of-Experts Transformer derived from the Grok 4 base, distilled to use about 40 percent fewer reasoning tokens per response. xAI ships it as two API model strings, grok-4-fast-reasoning and grok-4-fast-non-reasoning, rather than a single model with a toggle. It scores 35 on the Artificial Analysis Intelligence Index in reasoning mode, versus 33 for base Grok 4. The model targets large-context analysis and agentic tool-calling rather than raw chat quality. It was designed to sit below Grok 4 on cost while staying close on benchmark performance. xAI deprecated it on May 15, 2026 in favor of Grok 4.3.
How much does Grok 4 Fast cost per 1M tokens?
Grok 4 Fast costs $0.20 per 1 million input tokens and $0.50 per 1 million output tokens through the direct xAI API. Cached input tokens cost $0.05 per 1 million, a 75 percent discount off the standard input rate for repeat-context workloads. xAI applies a higher, unpublished rate once a single request crosses the 128,000 token threshold, a tiered-pricing pattern shared with base Grok 4. A 500,000-token codebase summary with a 20,000-token output costs roughly $0.11. A daily agentic coding loop pushing 2 million input and 200,000 output tokens runs about $0.50. A 1,000-turn support chatbot at 2,000 input and 500 output tokens per turn costs around $0.65. These rates were unchanged as of July 2026 despite the May 2026 deprecation notice. The model is proprietary and cannot be self-hosted, so there is no infrastructure cost alternative.
What is Grok 4 Fast's context window and max output?
Grok 4 Fast has a 2,000,000 token context window, one of the largest of any production model at its September 2025 launch, ahead of GPT-4.1's 1 million and Claude Sonnet 4.6's 200,000 token windows. Despite that huge input capacity, the xAI playground caps a single response at 16,000 output tokens, so the model is built for ingesting large inputs rather than generating very long outputs. xAI has not published independent needle-in-haystack recall figures specific to Grok 4 Fast, so recall quality above 1 million tokens is not independently verified on hokai. There is no separate extended-context tier; the same 2M window applies to both the reasoning and non-reasoning model IDs. Document handling works by passing the full text, image, or code directly in the prompt rather than chunking. xAI recommends the API directly, not the playground, for testing genuinely long-output tasks.
How does Grok 4 Fast compare on benchmarks vs Grok 4?
Grok 4 Fast Reasoning scores 35 on the Artificial Analysis Intelligence Index, edging out base Grok 4's score of 33 on the same composite, while the later Grok 4.1 Fast scores 39. xAI does not separately publish per-model GPQA Diamond or AIME 2025 scores for the Fast variant, stating only that it matches Grok 4's frontier benchmark results using about 40 percent fewer reasoning tokens. In a narrower, independently verified slice, the grok-4-fast-search variant ranked first in the LMArena Search Arena at 1163 Elo, 17 points ahead of OpenAI's o3-search. Base Grok 4 itself scored 88.9% on GPQA Diamond and up to 100% on AIME 2025 with tools, according to xAI's own reporting, but those specific figures have not been independently confirmed for the Fast variant. In practice this means Grok 4 Fast is priced and tuned to approximate Grok 4's reasoning quality at a fraction of the cost, not to exceed it. The absence of separately published GPQA and AIME numbers for the Fast model is itself notable: buyers evaluating raw reasoning quality should treat the Intelligence Index score as the more reliable comparison point.
Is Grok 4 Fast open source or proprietary?
Grok 4 Fast is proprietary and API-only; xAI has not released its weights. Access requires an API key from xAI's direct API at api.x.ai, using an OpenAI-compatible SDK surface, or through the OpenRouter aggregator at openrouter.ai/x-ai/grok-4-fast. Oracle Cloud briefly listed it in its Generative AI catalog but that listing is now marked deprecated as well. There is no confirmed availability through AWS Bedrock or Google Vertex AI. Commercial use is governed by xAI's standard API Terms of Service; there is no separate open-weights license or Hugging Face download. Because the model cannot be self-hosted, there are no VRAM or quantization requirements to consider. Teams needing an open-weights alternative with a similar context window should look elsewhere in the market rather than expect xAI to release these weights.
What modalities does Grok 4 Fast support?
Grok 4 Fast accepts text and image input and produces text output, plus native tool-call output for function calling and agentic workflows. It does not accept audio or video input and has no audio or video generation of its own. Through xAI's Agent Tools API, it can call live X search, general web search, and remote code execution mid-conversation, which is the main reason it gets used for agentic research and support automation rather than as a plain chatbot. Function calling and structured output are both supported natively through the API. There is no confirmed parallel tool-call limit published by xAI for this specific model. Compared to Grok 4.5, which added native video input, Grok 4 Fast is text-and-image only. Vision input covers standard image formats for document and chart reading, not video frames.
Does Grok 4 Fast train on user data?
xAI has not published a specific data retention or training-on-inputs policy for Grok 4 Fast that hokai can independently verify as of July 2026. The model's September 2025 model card describes xAI's general safety training approach, including refusals for CBRN-related requests, CSAM, and self-harm content, plus a system prompt layer with further mitigations, but it does not detail enterprise data retention terms in the sections reviewed. xAI has not published SOC 2, ISO 27001, HIPAA, or GDPR compliance statements specific to this model that could be confirmed here. Enterprise customers should request xAI's current data processing terms directly before sending sensitive data, rather than assume parity with competitors that publish detailed trust center documentation. Given the model is already deprecated, new enterprise commitments are better directed at Grok 4.3 in any case. No EU AI Act classification specific to this model was found in public documentation.
Who is Grok 4 Fast best for and who should avoid it?
Grok 4 Fast fits teams still running existing integrations that need large-context, low-cost inference and don't want to migrate immediately: its 2 million token window and $0.20/$0.50 per 1M token pricing suit large-document analysis, large-codebase review, and agentic search built on xAI's Agent Tools API. It is also a reasonable pick for cost-sensitive workloads where the Artificial Analysis Intelligence Index score of 35 is close enough to Grok 4's 33 to not matter much. It is the wrong choice for any brand-new long-term integration, since xAI deprecated it on May 15, 2026 and will fully retire the API on August 15, 2026, redirecting new traffic to Grok 4.3. Teams that need independently verified per-benchmark scores like GPQA or AIME for a specific model, rather than a composite index, should also look elsewhere, since xAI has not separately published those for the Fast variant. Latency-sensitive real-time applications may prefer the non-reasoning variant specifically, since reasoning mode adds meaningful response time. Anyone starting a new project in mid-to-late 2026 should evaluate Grok 4.3 or Grok 4.5 directly instead.