xAI releases improved Grok-2 model to all X users

• A faster, more capable Grok-2 can cite sources and produce more photorealistic images.

• It's more deeply integrated with the X platform.

• The "unfiltered" model can generate controversial answers and images.

X is rolling out the latest version of its Grok AI model for free to users across the social media platform.

The improved Grok-2 large language model can now search the web and produce highly realistic pictures.

X's AI company xAI says the model is three times faster than its predecessor and should provide more accurate results across multiple languages.

Grok-2 can answer queries, generate images and digest media from users. It now cites external websites and X posts in its replies.

It uses xAI's new Aurora model to generate images, all of which are watermarked with "Grok-2."

Aurora is an autoregressive image generation model intended to provide more realistic-looking pictures, as you can see from the image below.

All X users can access the new model by pressing the Grok icon on the left hand panel in X, or by heading directly to x.com/i/grok.

xAI is trying to make the LLM a standard feature of the X platform, and has added a Grok-powered "Explain this post" button to the top right-hand corner of all posts to encourage its use.

Although Grok-2 is freely available, Premium and Premium+ X accounts will get higher usage allowances and faster access to future updates. A beta version of the model has been available since August.

Developers can access Grok's improved capabilties directly via the the grok-2-1212 and grok-2-vision-1212 models in the xAI API.

Aurora is not yet available via the API, but xAI says it plans to publish the model soon.

Performance and pricing

Grok-2 scores reasonably well against its closest rivals according to instruction following benchmarking data released last week.

When it was first released in August, Grok-2 held its own on standard math, coding and accuracy benchmarks, beating OpenAI's GPT-4o on several scores.

OpenAI and Anthropic have both released new versions of their flagship LLMs in recent months, and it's not clear how well these compare to the current version of Grok-2.

Grok-2 is priced similarly to GPT-4o. It costs $2 per 10 million output tokens and $10 for 1 million output tokens, and new users currently get $25 free API credits.

GPT-4o costs $2.5 per million input tokens and $5 per million output tokens. That being said, heavy OpenAI users can slash their costs by using batch requests and caching, whereby bulk or similar queries are charged at half the regular cost.

Anthropic's rival Claude Sonnet-3.5 is the priciest of the three, costing $3 per million input tokens. Ouput tokens are $15 per million.

Again, heavy users can cut their costs with a caching discount.

Why "Grok"?

Named for a martian in the Robert A Heinlein novel "Stranger in a Strange Land," Grok is intended to portray a deep understanding of topics it discusses.

As the model itself told Indie Hackers:

"xAI chose this name for me because it reflects my purpose: to help users understand the universe and its many mysteries in a deep, insightful way. When you ask me questions, I aim to provide answers that not only inform but also resonate on a more intuitive level."

Grok is also sold as providing "unfiltered answers" alongside its reasoning, coding and visual processing skills.

Rivals like OpenAI place considerable emphasis on model safety to prevent the generation of harmful material, misinformation and deepfake content. Users of its new Sora video model cannot currently use their own images of people to create videos in an effort to prevent deepfakes.

Grok-2, on the other hand, has already been used to make a host of controversial images. When the preview model dropped in August, users flooded X with deliberately tasteless pictures they'd made with the LLM.

Given Aurora's capabilities, users can now generate much more realistic images. It doesn't appear any significantly tighter restrictions have been put in place.