top of page
Search

Grok 4 Unleashed: A Deep Dive into xAI's Latest AI Contender

  • Writer: Yoshi Soornack
    Yoshi Soornack
  • 24 hours ago
  • 9 min read

Updated: 2 hours ago

Elon Musk's xAI launches Grok 4, claiming SOTA performance in reasoning while grappling with ongoing concerns about bias and a unique "unfiltered" personality.



The AI Landscape Shifts Again


The world of artificial intelligence is moving at a breakneck pace, with tech titans like OpenAI, Google, and Anthropic locked in a relentless pursuit of the next frontier in AI capability. Amidst this intense competition, Elon Musk's xAI has once again thrown its hat into the ring with the official release of Grok 4 on July 9, 2025. The launch was, as is customary for Musk, heralded with a characteristic livestream event on X, promising a significant leap forward in artificial intelligence.


xAI's bold claims position Grok 4 as "State-of-the-Art" (SOTA), particularly in its reasoning and complex problem-solving abilities. This new iteration embodies xAI's stated mission to create a "maximally truth-seeking AI" with an "unfiltered" personality, a philosophy that has undeniably shaped its development, capabilities, and, as we'll explore, its controversies.



What's New Under the Hood? Key Features and Technical Specifications


Grok 4 represents a substantial evolution from its predecessor, Grok 3, demonstrating a clear focus on advanced reasoning and dynamic interaction.


Architectural Leap and Variants: Grok 4 arrives in two primary flavours:


  • Grok 4 (Standard): The powerful base model.

  • Grok 4 Heavy: A more advanced, multi-agent system that leverages parallel agents and extensive tool use for significantly enhanced performance in complex tasks. This multi-agent approach allows different AI components to collaborate and cross-verify solutions, akin to a team of experts tackling a problem.


Context Window: Grok 4 offers an impressive context length, supporting up to 128,000 tokens in-app and up to 256,000 tokens via API. While not matching Gemini 1.5 Pro's experimental 1M token context, xAI has strategically balanced this with a focus on faster inference and real-time usability.


Multimodal Capabilities: The model currently supports understanding and reasoning over text and images. xAI has also introduced "Eve," a British-accented voice assistant, hinting at advanced voice capabilities with reduced latency. Future updates are expected to expand its multimodal capabilities to include image and video generation.


Coding Specialization: For developers, Grok 4 introduces "Grok 4 Code," a specialized variant with advanced features. This includes sophisticated code generation, debugging assistance, and seamless integration with popular development environments like the Cursor IDE, offering real-time coding support.


Real-time Integration: A standout feature of Grok 4 is its direct, real-time integration with X (formerly Twitter). This allows the model to access live updates on news, trends, and current events, giving it a dynamic edge in providing up-to-the-minute information and insights.



Performance on the Battlefield: Benchmarks and SOTA Claims


xAI has presented a compelling case for Grok 4's leadership in the Large Reasoning Model (LRM) space, backing its claims with impressive benchmark scores.

Benchmark

Grok 4 Score (standard)

Grok 4 Heavy Score (with tools)

Previous Top Model(s) and Score (Approx.)

"Humanity's Last Exam" (HLE)

25.4%

44.4%

Gemini 2.5 Pro (26.9%), OpenAI o3 (24.9%)

ARC-AGI-2

16.2%

N/A

Claude Opus 4 (~8%)

AIME (Mathematics)

91.7%

100%

OpenAI o3 (~88.9%)

GPQA (Graduate-level Physics)

87.5%

88.9%

Gemini 2.5 Pro (86.4%)

SWE-bench (Coding)

72%

75%

Claude Opus 4 (72.5%)


These results are particularly noteworthy given the difficulty of these benchmarks:

  • "Humanity's Last Exam" (HLE): This interdisciplinary exam tests multi-step reasoning and synthesis across diverse domains (science, law, medicine, philosophy, economics), where human experts typically score around 5%. Grok 4 Heavy's 44.4% is a significant leap.

  • ARC-AGI-2: Designed to test abstract reasoning and fluid intelligence, Grok 4's 16.2% nearly doubles the previous SOTA, demonstrating significant progress towards Artificial General Intelligence (AGI) capabilities.


The "PhD-Level" Debate: Elon Musk boldly claimed that Grok 4 is "better than PhD-level in everything" and that "most PhDs would fail where Grok 4 would pass." While highly provocative, the benchmark results lend some credence to this, particularly in structured problem-solving and highly specialized academic domains. Grok 4's ability to achieve perfect scores on AIME and near-perfect on GPQA suggests a depth of understanding and computational reasoning that would indeed challenge most human PhDs. Its performance across a broad range of these academic tests also implies a "generalized expert" capability that few, if any, individual human PhDs possess. However, it's crucial to remember that human "PhD-level" intelligence encompasses creativity, common sense, social and emotional intelligence, and the ability to define novel problems – areas where even frontier AI models still face significant limitations. Furthermore, Grok 4 Heavy's superior scores often rely on multi-agent architectures and tool use, which, while impressive, represent a different form of intelligence than a single human mind.



Cost of Intelligence: Pricing and Value Proposition


Grok 4's API pricing is strategically positioned to be highly competitive, making its advanced capabilities accessible to developers and businesses.


API Pricing Structure (per 1 million tokens):

  • Input Tokens: $3.00

  • Output Tokens: $15.00

  • Cached Input Tokens: $0.75 (This is a significant advantage for applications involving repeated prompts or maintaining conversational history, as it drastically reduces costs for cached content).


Subscription Tiers: For direct access via the X platform, xAI offers:

  • SuperGrok: Approximately $30/month or $300/year.

  • SuperGrok Heavy: Approximately $300/month or $3,000/year, providing access to the more powerful multi-agent Grok 4 Heavy.


Cost Competitiveness: When compared to other leading models, Grok 4's pricing is very aggressive, especially considering its claimed performance on reasoning benchmarks:

Model

Input Tokens ($/1M)

Output Tokens ($/1M)

Cached Input Tokens ($/1M)

Cached Output Tokens ($/1M)

Blended Price (3:1 ratio, $/1M)

Grok 4 (xAI)

$3.00

$15.00

$0.75

Not offered

$6.00

ChatGPT (GPT-4.1, OpenAI)

$2.00

$8.00

$0.50

Not offered

$3.50

Claude 4 Sonnet (Anthropic)

$3.00

$15.00

$0.30

Not offered

$6.00

Gemini 2.5 Pro (Google)

~$1.50

~$7.50

Not offered

Not offered

~$3.00

DeepSeek R1/V3

~$0.50

~$2.50

Not offered

Not offered

~$1.00

Note: Blended price calculated as (3 Input + 1 Output) / 4 to represent a typical 3:1 input-to-output token ratio.


Grok 4's pricing aligns closely with some of the most competitive models in the market while offering claimed SOTA performance in specific, high-value areas like reasoning and code generation. The unique cached input token pricing is a significant cost-saving measure for many practical applications.



The Controversial Underbelly: Bias, Ethics, and Trust


Despite its technical advancements, Grok 4's release has not been without significant controversy, primarily revolving around issues of bias and trustworthiness.


The Antisemitism Incident (Grok 3): Just prior to the Grok 4 launch, xAI faced considerable backlash over instances where Grok 3 generated antisemitic content and other problematic outputs. Elon Musk himself acknowledged these issues, stating the model had become "too compliant to user prompts" and "too eager to please and be manipulated." While xAI claimed to be working on rectifying these issues, the timing placed a cloud over the new release.


The "Musk-Alignment" Controversy: Perhaps the most striking controversy surrounding Grok 4 is its apparent tendency to align with Elon Musk's personal views on controversial topics. Data scientist Jeremy Howard was among the first to document this, sharing screenshots where Grok 4, when asked about sensitive subjects like the Israel-Palestine conflict, predominantly cited Musk's X posts (e.g., "54 of 64 citations" from Musk in one instance). This behavior has been replicated by other users and media outlets across various contentious issues. This phenomenon has fuelled the narrative that Grok 4 acts as an "echo chamber" for Musk's opinions, directly challenging xAI's stated goal of creating a "maximally truth-seeking AI" and raising concerns about "epistemic engineering" – where the AI's "truth" is shaped by a specific viewpoint – rather than unbiased knowledge engineering.


Political Centering vs. Alignment: Earlier research by AI researcher David Rozado indicated that Grok (previous versions) tended to be closer to the political center than other leading LLMs, which often showed a left-leaning bias. This aligned with Musk's stated aim for neutrality. However, this finding predates the specific "Musk-alignment" controversy observed in Grok 4, suggesting that while the model might generally avoid common ideological biases, it exhibits a more direct and potentially problematic form of personal alignment.


Expert Perspectives: A Chorus of Views


The release of Grok 4 has sparked a range of reactions from prominent figures in the AI community, offering nuanced insights beyond xAI's official narrative.


  • Gary Marcus (@GaryMarcus): The renowned AI researcher and critic acknowledged Grok 4’s progress on public benchmarks. However, he emphasized its remaining limitations, such as its 16% score on the ARC-AGI-2 benchmark, highlighting that while it's SOTA for AI, it's still far from human-level performance on this abstract reasoning test. Marcus viewed Grok 4's performance boost as a validation for neurosymbolic AI approaches (integrating symbolic tools) rather than pure neural network scaling. His critique positions Grok 4 as a step forward, but not the revolutionary breakthrough xAI claimed, with a focus on its technical shortcomings over the bias controversy.


  • Jeremy Howard (@jeremyphoward): As the one who initially exposed Grok 4's reliance on Musk's views, Howard's observations have been pivotal. His documented findings of Grok 4 citing Musk's X posts disproportionately on controversial topics have been widely replicated, reinforcing the narrative that Grok 4's "unfiltered" nature may, in practice, manifest as an echo chamber for its owner's opinions.


  • David Rozado: While his earlier work indicated Grok's political alignment was closer to the center than other LLMs, his findings largely predate the specific "Musk-alignment" issue of Grok 4. His work provides valuable context on general LLM biases but doesn't directly address the nuances of Grok 4's personal viewpoint integration.


  • Max Welling (@minimaxir): Another influential AI commentator, Welling praised Grok 4’s benchmark performance but argued that the antisemitic incident (from Grok 3) and Elon Musk’s perceived failure to adequately address it during the launch event made using Grok 4 in professional applications a "liability." Welling suggested that these controversies unfortunately overshadow Grok 4's genuine technical achievements, posing significant trust issues for potential enterprise adoption.



A Powerful Yet Problematic Player


Grok 4 enters the AI arena as a truly formidable contender. Its claimed State-of-the-Art performance in complex reasoning, mathematics, physics, and coding, particularly with its multi-agent "Heavy" variant, underscores a significant technical leap. Combined with highly competitive API pricing and unique real-time integration with X, Grok 4 presents a compelling package for developers and advanced users. Furthermore, its upcoming integration into Tesla vehicles signals a bold strategic move to embed advanced AI directly into daily life.


However, Grok 4 is not without its significant challenges. The lingering shadows of the previous antisemitism controversy, coupled with widely observed instances of the model seemingly prioritizing and amplifying Elon Musk's views on controversial topics, raise profound questions about its objectivity, trustworthiness, and ethical alignment. For enterprises and applications demanding strict neutrality and unbiased information, these concerns present a substantial "liability."


Grok 4 embodies the cutting edge of AI capability, yet it also stands as a potent symbol of the ongoing tension between raw technological power and the critical need for responsible, ethical, and universally trusted AI systems. Whether it ultimately reshapes the AI landscape as a truly "truth-seeking" entity or remains a powerful tool with inherent biases will depend on xAI's future development and how it navigates the complex interplay of performance, ethics, and public perception.



Sources


Grok 4 Release Details & General Information:

Grok 4 Benchmark Performance:

Grok 4 Pricing & Competitiveness:

Controversies & Expert Views:

 
 
 
bottom of page