Founded just two years ago in Mar 2023, Elon Musk’s xAI has had some work to do to catch up with OpenAI, which was founded in Dec 2015 (originally backed by Musk and others) and had a 7-year headstart. Grok 3, which was released this past Monday, is now making waves with knowledgeable industry watchers such as Andrej Karpathy placing it among the state-of-the-art models (although it doesn’t outperform at every task). As of this writing, Grok 3 is currently placed #1 on the Chatbot Arena leaderboard (which pits models head-to-head to achieve a crowdsourced set of evaluations).

While xAI’s earlier Grok-2 was able to achieve a #4 ranking on the Chatbot Arena leaderboard, Grok 3 is the first of xAI’s models to have led the board, albeit by a slim margin. Grok 3 is currently ahead of Google’s Gemini 2.0 Flash Thinking Experimental and Gemini 2.0 Pro Experimental, OpenAI’s 4o, and DeepSeek’s R1 reasoning model on the board.

Chatbot Arena’s rankings, however, do not include OpenAI’s o3 reasoning model, which outperforms Grok 3 on math and coding based on self-reported scores. (o3 has not been released yet, and will only be released as part of GPT-5 rather than as a standalone model.) Industry watchers are still looking for independent verification of Grok 3’s performance. Some have noted that Grok 3 performs better in Chatbot Arena than in the real world, where it can fail sometimes on even simple questions.

Grok 3 is actually a family of models, and two will be released initially: Grok 3 and Grok 3 Mini. xAI describes Grok 3 as “blending strong reasoning with extensive pretraining knowledge” with capabilities that were “refined through large scale reinforcement learning.” It offers 3 different modes – “Standard,” “Think” (multi-step reasoning), and “Big Brain” (complex, slower reasoning). Like OpenAI’s o3, Grok 3 uses test-time computing for its reasoning. It also offers a context window of 1M tokens – one of the larger context windows among the major LLMs. Grok 3 Mini is similar but provides more “cost-efficient reasoning.” Both models are still in training. The version of Grok 3 that topped the leaderboards, xAI says, was an early version codenamed chocolate.

The Grok reasoning model also powers xAI’s new DeepSearch tool – a tool akin to the “deep research” tools recently introduced by OpenAI, Google, Perplexity, and Hugging Face. According to Karpathy, xAI’s version appears to be on par with Perplexity’s Deep Research but not at the level of OpenAI’s Deep Research.

X Premium and Premium+ users have early access to Grok 3, and xAI enterprise API users will get access within a few weeks. xAI revealed a new subscription called SuperGrok for the standalone version of Grok via its app and website. SuperGrok will include higher limits for Grok and DeepSearch queries and image generation. The Grok app will soon get native voice mode (and audio-to-text transcription), and Musk has said the Grok.com website will have the “latest and most advanced version” of Grok. The SuperGrok subscription is priced at $30/month ($300/year).

The Grok family has been noted to have fewer guardrails than other state-of-the-art models. Its less filtered results have, at times, been described as “humorous,” “edgy” and anti-woke. In response to a question about The Information, it responded: “The Information, like most legacy media, is garbage.” At one point, Musk suggested that Grok might get an “unhinged mode” that would produce results “intended to be objectionable, inappropriate, and offensive.” Looser guardrails also means Grok can be more readily “jailbroken” or otherwise hacked. This also may be making Grok more competitive in Chatbot Arena, which favors models that are more willing to respond to requests.

The history of xAI can be traced back to OpenAI. OpenAI Inc – also known as OpenAI Nonprofit – was founded in Dec 2015 as a nonprofit research institution by Musk and Sam Altman (at the time, the president of Y Combinator). OpenAI’s mission was to ensure that “artificial general intelligence” (AGI) benefits all of humanity. The nonprofit was backed by $1B+ in notional commitments from Musk – who ended up putting in $44M – as well as Altman, Greg Brockman, Reid Hoffman, Jessica Livingston, Peter Thiel, Amazon (AWS), Infosys, and YC Research. A TechCrunch analysis added up $133M in total actual donations to the nonprofit from 2015-2021, which roughly matches what OpenAI has said itself.

Around 2017, the OpenAI team began realizing the sheer amount of capital they would need to advance their mission, and discussed a transition to a for-profit. Based on contemporaneous emails and documents from Musk’s later lawsuit against Altman, there were disagreements on direction and control over OpenAI. Both Musk and Altman appear to have sought the CEO role. For the time being, OpenAI stayed a nonprofit under Altman’s leadership. Elon Musk left the board in Feb 2018 on reasonably good terms, both due to the disagreements and to avoid conflicts with Tesla. According to OpenAI, Musk planned to build an AGI rival within Tesla.

With OpenAI’s ascendance came growing animosity from Musk, who became critical of its closed-source, profit-oriented approach to AI development alongside Microsoft’s growing influence. It culminated in the launch of his xAI (X.AI Corp) in Mar 2023. xAI was structured as a benefit corporation, with the mission of “understand[ing] the universe.” It has become closely affiliated with Musk-owned X (formerly Twitter), which is now a significant channel for xAI products.

Musk has raised $12B+ from investors such as a16z, Sequoia, Fidelity, Lightspeed, Valor Equity Partners, Nvidia, AMD, and Saudi and Qatari entities. xAI’s most recent round was a $6B Series C at a $51B valuation in Dec 2024. As of last week, Musk was reportedly looking for another $10B at a $75B valuation.

xAI’s team today has grown to about 100 noncontract employees, including staff with prior experience at OpenAI, Google/DeepMind, Microsoft, Meta, Tesla, and the University of Toronto. xAI also has 900+ “AI tutors” or data annotators on contracts who help train Grok, with plans to add “thousands more” this year.

In the two years since its founding, xAI has released Grok-0 (Aug 2023), Grok-1 and a Grok chatbot (Nov 2023), its PromptIDE environment for prompt engineering (Nov 2023), the first Grok API (Nov 2023), Grok for X Premium+ users ($40/month) (Dec 2023), Grok-1 open-source release (Mar 2024), Grok-1.5 (Mar 2024), Grok-1.5V with vision (Apr 2024), Grok-1.5 for X Premium users ($8/month) (May 2024), Grok-2 (Aug 2024), xAI enterprise API with access to all its models (Nov 2024), Grok with Aurora image generation on X (Dec 2024), Grok-2 for X free users (Dec 2024) with higher limits for paid subscribers, and a standalone Grok app (Jan 2025).

Notably, Grok 3 is the first of xAI’s models to be trained on its new Memphis-based Colossus supercluster. xAI calls it the “world's largest AI supercomputer” with 100K Nvidia Hopper GPUs or “10x the compute of previous state-of-the-art models.” According to xAI, Colossus was “fully operational in 122 days [starting in Apr 2024] and started running workloads just 19 days after the first servers were delivered.” xAI then doubled Colossus’ size to 200K Nvidia Hopper GPUs (using Nvidia’s Spectrum-X networking) in 92 days. It is already working on its next cluster, which will use 5x the power or about 1.2 GW and would make for 1M+ total GPUs in the supercluster. xAI is reportedly in talks on a $5B+ deal to buy Dell servers with Nvidia chips.

Access to compute is clearly one of Grok’s major advantages, spurring its progress. Some are asking, however, whether the slim margins by which Grok is outperforming its next rivals are worth the amount of compute used to obtain those gains. Like other AI players, xAI is working to make its models faster and more efficient. These efforts have already allowed xAI to slash its API pricing by 33-60% in Dec 2024. Some have speculated that Grok 3 may be using a very sparse mixture-of-experts (MoE) architecture, akin to how DeepSeek achieves its efficiencies.

Musk plans to open-source older versions of Grok as soon as the newest model is mature and stable. He says this could happen “within a few months” for Grok 3. Given the pace at which xAI is moving, it’s not much of a loss to open-source Grok 2 (its team has described Grok 2 as a “toy” vs. Grok 3). Opening up Grok-2 could draw users, developers, and scientists into its ecosystem, and align with Musk’s criticism of OpenAI’s lack of openness, without significantly cannibalizing Grok 3.

xAI continues to be shaped by its rivalry with OpenAI and the vitriol between Musk and Altman. OpenAI has asked its investors not to invest in its rivals including xAI. The ask for exclusivity was nonbinding – ARK Venture Fund, for instance, was already invested in xAI and Anthropic, and Fidelity has continued to invest in xAI. Altman also surprised Musk with last month’s announcement of OpenAI’s $500B Stargate infrastructure investment (with SoftBank and Oracle) as a signature Trump initiative.

Musk, in turn, initiated a lawsuit against Altman in Aug 2024, that has been highly revelatory in the details exposed regarding OpenAI’s earlier internal discussions. More recently, Musk put in an unsolicited $97B bid for OpenAI, throwing a wrench into its for-profit transformation, which is in full swing. The bid was rejected by Altman, not unsurprisingly. Nevertheless, it puts outside pressure on OpenAI’s valuation of its assets, the full value of which needs to stay with the nonprofit entity (i.e. the assets would probably need to be sold to the for-profit). Altman’s comment that Musk was “trying to slow us down” was probably not too far from the truth.

xAI has now proven itself to be a major AI player. Its progress has been heavily reliant on Musk’s ability to recruit top-tier talent, raise large funding rounds to pay for compute, and move rapidly to operationalize the capital. While Grok 3 was slightly behind its target release date (end of 2024), in general xAI has caught up impressively fast. Grok probably remains behind OpenAI’s o3 in terms of performance but that situation could change quickly. As the AI players move towards more unified consumer/developer experiences, xAI’s association with X – which has 600M+ users – may prove to be a distinctive advantage.

Feb 7 2025 (3 Shifts): Distillation and AI economics

Jan 10 2025 (3 Shifts): GPT-5's troubles and what's next for AI data