Musk’s AI company, now operating as SpaceXAI after previously being known as xAI, has moved quickly this year. Three Grok releases landed in under three months, each aimed squarely at developers and agent workloads.

Grok 4.5 (July 8)

  • Positioned as a workhorse for coding, app-building, office work, research and writing.
  • Cost: two dollars for every million tokens sent in, six dollars for every million generated.
  • Musk called it an “Opus-class model”, and the company claimed roughly twice the token efficiency of other leading models. TechCrunch noted its published benchmarks were competitive but just short of best-in-class.
  • It launched through Grok Build, the Cursor editor and the company’s API console, but was not available in the EU at launch.

Grok 4.6 (August 12)

The next release focused on long-running agents and was made available through the company’s API and third-party platforms, including GitHub Copilot and Amazon Bedrock.

Grok 4.7 (September 21)

Grok 4.7 kept the same pricing as 4.6. Its training used data from Cursor, which the company acquired this year, and observers noted it consumed more tokens on the same tasks.

The bigger picture

The strategy is clear: undercut rivals on price, ship often, and get Grok into the tools developers already use. The open question is quality versus the very top models. No release date has been announced for Grok 5.

A short history of Grok

Musk founded xAI in 2023, and the first Grok model appeared that November, available to subscribers on X. Early versions were known mainly for a sarcastic personality and access to posts on X. Over the following two years, the company built one of the largest AI computing clusters in the world, released models that competed near the top of public benchmarks, and expanded Grok to its own apps and an API for developers. This year’s rapid releases continue that push from chatbot novelty to serious developer tool.

Why distribution matters

Developers rarely switch models just because a new one appears. They use whatever is built into their editor, their cloud platform or their coding agent. Getting Grok into GitHub Copilot, Amazon Bedrock and the Cursor editor puts it in front of developers who would never sign up for a separate service. Owning Cursor also gives the company a direct view of how people actually write code with AI, which is valuable training data, as the 4.7 release showed.

Price per token vs price per task

Headline prices can mislead. A model that charges less per million tokens but uses more tokens to finish the same job can cost the same or more in practice. Reports that Grok 4.7 consumed more tokens than 4.6 on the same tasks are a good reminder: the number that matters to a business is the total cost of a finished, correct result.

To compare fairly, run the same set of real tasks through each model and record three things: whether the result was correct, how many tokens it used and how long it took.

What we think

Three releases in under three months shows how fast the model race moves, and low prices make Grok easy to try. The flip side is a moving target for anyone building on it, since behavior and token use changed between versions. If you build on Grok, pin a specific model version, measure cost per task rather than price per token, and test again before every upgrade.