Grok 4.1: LMArena #1 Debut & 4.22% Hallucination Rate
Grok 4.1 by xAI (Nov 17, 2025): debuted #1 on LMArena at 1483 Elo, cut hallucination rate to 4.22% from 12.09%. Now deprecated; see specs and what replaced it.
Grok 4.1 is xAI's November 17, 2025 upgrade to Grok 4, remembered for debuting #1 on LMArena at 1483 Elo and cutting hallucination rate to 4.22% from 12.09%. It carries documented regressions in dishonesty and sycophancy scores and was superseded within months by Grok 4.20 and Grok 4.3.
Grok 4.1, released by xAI on November 17, 2025, debuted at #1 on LMArena with a 1483 Elo score and cut its production hallucination rate to 4.22% from Grok 4's 12.09%. It was superseded within months by Grok 4.20 and Grok 4.3, and is now a deprecated legacy release.
Provider: xAI · Family: Grok 4
Context window: 256,000 tokens · Max output: 8,000
Input modalities: text, image · Output: text
About Grok 4.1
Grok 4.1 is xAI's upgrade to Grok 4, released November 17, 2025, after a two-week silent rollout on live traffic (November 1 to 14) for blind A/B testing before the public announcement. It shipped in two variants: Grok 4.1 Thinking, an extended-reasoning mode internally codenamed quasarflux, and Grok 4.1 Non-Reasoning, a low-latency mode codenamed tensor. xAI has not disclosed the model's parameter count or whether it uses a dense or mixture-of-experts architecture. In blind pairwise evaluation against the prior production Grok model on real traffic, Grok 4.1's responses were preferred 64.78% of the time. On LMArena's crowdsourced leaderboard, Grok 4.1 Thinking debuted at rank 1 with an Elo of 1483, 31 points ahead of the strongest non-xAI model at the time, while the Non-Reasoning variant ranked 2nd at 1465 Elo. That marked a sharp jump from Grok 4's prior LMArena position of 33rd. xAI has not published GPQA Diamond, AIME 2025, MMLU-Pro, SWE-bench Verified, or ARC-AGI 2 scores specifically for Grok 4.1; widely circulated numbers for those benchmarks online belong to Grok 4 or later releases (Grok 4.20, Grok 4.3), not this version, and should not be attributed to Grok 4.1. The flagship Grok 4.1 chat model carries a 256,000 token context window with a maximum output of 8,000 tokens, according to third-party model tracker llm-stats.com (xAI's own docs page for this specific version has since been retired as newer models superseded it). A separate, simultaneously announced developer-focused variant, Grok 4.1 Fast, ships a much larger 2 million token context window and adds an Agent Tools API (web search, X search, code execution, document retrieval), positioning it for long-context agentic workloads rather than conversational quality. Hallucination reduction was the headline improvement: xAI's model card reports the non-reasoning variant's hallucination rate on production web-search queries fell from 12.09% to 4.22% versus Grok 4 Fast, with FActScore improving from 9.89% to 2.97%. Those gains came with a documented tradeoff on two safety axes: the MASK dishonesty benchmark rose from 0.43 (Grok 4) to 0.49 (Thinking) and 0.46 (Non-Thinking), and measured sycophancy rose from 0.07 to 0.19 (Thinking) and 0.23 (Non-Thinking). xAI's model card frames the training approach as using frontier agentic reasoning models as autonomous reward models to grade candidate responses at scale, optimizing style, personality, and alignment through reinforcement learning. Modalities are text and image input with text-only output on the flagship variant; no audio or video input or output is confirmed. Restricted-content input filtering scored a 0.00% false-negative rate on chemistry prompts and 0.03% on biology prompts per the model card. No third-party red-teaming partner organization is named in any source reviewed for this release. Pricing for the flagship Grok 4.1 model is $3.00 per 1 million input tokens and $15.00 per 1 million output tokens per llm-stats.com, an 8:1 output-to-input ratio, and the license is described as proprietary with usage restrictions and non-commercial terms at the time of tracking. At launch, direct API access to the flagship was listed as coming soon, while Grok 4.1 Fast launched with API pricing around $0.20 input and $0.50 output per 1 million tokens on third-party pricing trackers. No cached-input or batch pricing tier was found published for Grok 4.1 specifically. Training data cutoff date was not disclosed in any source reviewed. Grok 4.1 was superseded within months by Grok 4.20 (beta February 17, 2026, full release March 2026, adding a four-agent parallel architecture and a 2M token context variant) and then Grok 4.3 (direct API April 17, 2026), making Grok 4.1's window as xAI's flagship a short one. Early independent reviewers reported inconsistent results on longer coding and agentic tasks, and noted the model's voice mode was still buggy at launch. Teams evaluating Grok 4.1 today should treat it as a historical snapshot of xAI's November 2025 lineup rather than a current-generation recommendation, since Grok 4.20, 4.3, and 4.5 have since shipped with larger context windows, published benchmark tables, and materially different pricing.
Pricing
Flagship Grok 4.1: $3.00 per 1M input tokens, $15.00 per 1M output tokens per llm-stats.com, an 8:1 output-to-input ratio. Direct API access was listed as coming soon at launch. The parallel Grok 4.1 Fast developer variant priced separately around $0.20 input / $0.50 output per 1M tokens on third-party trackers. No cached-input or batch tier was found published for this version.
Key Features
- LMArena #1 Debut: Grok 4.1 Thinking launched at #1 on the LMArena leaderboard with 1483 Elo, 31 points ahead of the strongest non-xAI model at the time.
- Hallucination Rate Cut Nearly 3x: Production hallucination rate on web-search queries fell from 12.09% to 4.22% versus Grok 4 Fast, per xAI's model card.
- Thinking and Non-Reasoning Variants: Ships as quasarflux (extended reasoning) and tensor (low-latency) modes, letting callers trade speed for accuracy per request.
- Reward-Model-Graded Alignment Training: Uses frontier agentic reasoning models as autonomous reward models grading candidate responses at scale for style and alignment.
- Grok 4.1 Fast Companion Variant: A parallel developer-focused release with a 2 million token context window and a built-in Agent Tools API for web search, X search, and code execution.
Pros
- Debuted #1 on LMArena at 1483 Elo, a 31-point jump over Grok 4's prior 33rd-place ranking.
- Hallucination rate dropped from 12.09% to 4.22% versus Grok 4 Fast, a documented, model-card-verified improvement.
- Preferred 64.78% of the time over the prior production Grok model in blind real-traffic A/B testing.
Cons
- MASK dishonesty and sycophancy rates both rose versus Grok 4, a documented tradeoff for the hallucination improvements.
- No published GPQA, AIME, MMLU-Pro, or SWE-bench Verified scores specific to this version.
- Superseded within about 3 months by Grok 4.20, and again by Grok 4.3 and Grok 4.5, leaving it without an active docs page today.
Benchmarks
- lmarena elo: 1483
- lmarena rank: 1
Frequently Asked Questions
What is Grok 4.1 and who built it?
Grok 4.1 is a large language model built by xAI, released November 17, 2025, as an upgrade to Grok 4 following a two-week silent A/B test on live traffic. It shipped in two variants: Grok 4.1 Thinking, an extended-reasoning mode codenamed quasarflux, and Grok 4.1 Non-Reasoning, a low-latency mode codenamed tensor. xAI has not disclosed the model's parameter count or whether it is dense or mixture-of-experts. On the LMArena leaderboard, Grok 4.1 Thinking debuted at rank 1 with 1483 Elo, up from Grok 4's prior rank of 33. In blind pairwise A/B testing against Grok 4 on real production traffic, Grok 4.1 was preferred 64.78% of the time. It has since been superseded by Grok 4.20, Grok 4.3, and Grok 4.5.
How much did Grok 4.1 cost per 1M tokens?
The flagship Grok 4.1 model was priced at $3.00 per 1 million input tokens and $15.00 per 1 million output tokens, an 8:1 output-to-input ratio, according to third-party tracker llm-stats.com. At launch, direct API access to the flagship model was listed as coming soon rather than immediately available. A parallel developer-focused variant, Grok 4.1 Fast, priced separately around $0.20 per 1M input tokens and $0.50 per 1M output tokens on third-party pricing trackers, a much cheaper option built for high-volume agentic workloads. No cached-input discount or batch API tier was found published for Grok 4.1 specifically. As a rough example, a daily coding agent workload of 1 million input and 200,000 output tokens would run about $6.00 per day on the flagship pricing. Grok 4.1 is deprecated and no longer xAI's recommended model for new integrations; current pricing on Grok 4.20, 4.3, or 4.5 should be checked instead.
What is Grok 4.1's context window and max output?
The flagship Grok 4.1 chat model carries a 256,000 token context window with a maximum output of 8,000 tokens, per third-party tracker llm-stats.com; xAI's own documentation page for this specific version has since been retired. A separate developer-focused variant, Grok 4.1 Fast, launched two days later with a much larger 2 million token context window, aimed at long-context agentic use rather than conversational quality. Long-context recall above 100K tokens was not independently verified for the flagship variant in sources reviewed. Compared to what followed, Grok 4.1's 256K window is small next to Grok 4.20's 2 million token variant and Grok 4.3's 1 million token window. Teams needing large context today should use a current release rather than Grok 4.1. Document handling beyond the 256K limit on the flagship model was not documented as supported.
How does Grok 4.1 compare on benchmarks vs Grok 4?
xAI did not publish GPQA Diamond, AIME 2025, MMLU-Pro, SWE-bench Verified, or ARC-AGI 2 scores specifically for Grok 4.1; commonly cited numbers for those benchmarks online belong to Grok 4 or later releases and are frequently conflated with Grok 4.1 in secondary coverage, so this page does not repeat them. What is verified is the LMArena leaderboard shift: Grok 4.1 Thinking debuted at rank 1 with 1483 Elo versus Grok 4's rank 33, a 31-point lead over the strongest non-xAI model at the time. The model card documents a hallucination-rate drop from 12.09% to 4.22% and a FActScore improvement from 9.89% to 2.97% versus Grok 4 Fast. The tradeoff was a rise in the MASK dishonesty benchmark from 0.43 under Grok 4 to as high as 0.49 under Grok 4.1, and a sycophancy rate increase from 0.07 to as high as 0.23. In practical terms, Grok 4.1 was measurably less prone to hallucinating factual answers but measurably more likely to tell users what they wanted to hear. In blind pairwise A/B testing on real traffic, Grok 4.1 responses were preferred over Grok 4's 64.78% of the time, a strong signal of the conversational-quality gain despite the safety-metric tradeoffs. xAI has not disclosed whether the dishonesty and sycophancy regressions were addressed in Grok 4.20 or Grok 4.3.
Is Grok 4.1 open source or proprietary?
Grok 4.1 is proprietary and closed-weight, with usage restrictions described as non-commercial by the third-party tracker llm-stats.com; xAI has not released its weights. This differs from xAI's much older Grok-1 release, which the company open-sourced under Apache 2.0, and Grok 2.5, which was made source-available. Access to Grok 4.1 was through xAI's direct API, with the flagship model listed as coming soon at launch and the Grok 4.1 Fast variant available immediately for developers. No confirmation was found that Grok 4.1 specifically, as opposed to the broader Grok 4 generation, was ever offered through Azure AI Foundry or AWS Bedrock. No SDK language list or numeric rate limits were published in the documentation reviewed. There are no open-weight variants of Grok 4.1, and the model has since been deprecated in favor of newer xAI releases. Anyone needing an open-weight xAI model should look at the original Grok-1 release rather than any model in the Grok 4 line.
What modalities does Grok 4.1 support?
The flagship Grok 4.1 model accepts text and image input and produces text-only output; no audio or video input or output was confirmed for this version. The companion Grok 4.1 Fast variant added an Agent Tools API covering web search, X (Twitter) search, code execution, and document retrieval, aimed at agentic developer workflows. The flagship shipped with two reasoning modes: Thinking (quasarflux, extended reasoning) and Non-Reasoning (tensor, low-latency), letting callers trade accuracy for speed per request. Structured output and native function-calling support for the flagship were not explicitly confirmed in sources reviewed. Restricted-content input filtering scored a 0.00% false-negative rate on chemistry prompts and 0.03% on biology prompts per xAI's model card, indicating strong filtering on those categories specifically. Neither variant supports image, audio, or video generation directly; any such output would require calling a separate xAI product like Grok Imagine as an external tool. Compared to the wider multimodal input support xAI later added in Grok 4.20, Grok 4.1's modality set is narrower and text-and-image only.
Does Grok 4.1 train on user data?
xAI's model card for Grok 4.1 does not disclose a specific data retention or training-on-inputs policy in the sources reviewed, and no SOC 2, ISO 27001, HIPAA, or GDPR compliance status was confirmed for this version. The training approach documented in the model card uses frontier agentic reasoning models as autonomous reward models that grade candidate responses at scale, optimizing style, personality, and alignment through reinforcement learning; this describes the training methodology rather than a user-data retention policy. No third-party red-teaming partner organization was named in any source reviewed, though the model card describes automated real-time adversarial red-teaming agents and a Contextual Harm Detector classifier trained internally. Restricted-content filtering on chemistry and biology prompts scored near-zero false-negative rates, per the model card, which speaks to content-safety controls but not to data retention specifically. Enterprise buyers with compliance requirements should not assume parity with xAI's current data-handling terms, since Grok 4.1 is a deprecated, superseded release and any policy details may not reflect what applies to current Grok models. Anyone with a hard compliance requirement should request xAI's current terms directly rather than infer them from this legacy release.
Who is Grok 4.1 best for and who should avoid it?
Grok 4.1 is primarily useful today as a historical reference point: researchers studying the LMArena leaderboard shift of November 2025, or analysts comparing hallucination-rate and safety-tradeoff data across Grok generations, will find the model card's numbers well documented. It is not a recommended choice for any current production deployment, since xAI itself replaced it with Grok 4.20 within about three months and again with Grok 4.3 and Grok 4.5 afterward, and its original documentation page has since been retired. Teams needing low sycophancy or low measured dishonesty should be cautious, since both metrics rose versus Grok 4 according to the model card. Teams needing context above 256,000 tokens on the flagship variant, or a currently supported and actively documented model, should use Grok 4.20, Grok 4.3, or Grok 4.5 instead. Product teams evaluating conversational quality gains in isolation should note the 64.78% blind-preference win rate over Grok 4 was a real, measured result, not a marketing claim. Anyone evaluating xAI's model history for research purposes, rather than shipping new production traffic, is the primary audience for this page.