Inside Elon Musk’s Grok‑4 Launch: Breakthrough AI Capabilities and Pricing

Elon Musk announced the release of Grok‑4, a new AI reasoning model that requires a subscription starting at $30 per month.

Grok Launch Overview

The official website is https://grok.com/ and the event video is linked in the source.

Model Introduction (5:22‑7:51)

Musk described Grok‑4 as "the smartest AI in the world" with reasoning abilities that achieve near‑perfect scores on high‑difficulty exams such as the SAT and GRE.

Exponential Performance Growth

Training compute increased roughly tenfold per generation, resulting in a hundredfold total increase from Grok‑2 to Grok‑4, with additional focus on inference (RL) compute to enhance problem‑solving.

Benchmark Results: "Humanity's Last Exam" and Others

The model was evaluated on a 2,500‑question benchmark covering hundreds of subjects, outperforming competitors including Claude Opus 4 and Gemini 2.5 Pro on tests like GPQA, AIME 25, and HMMT.

Live Demonstrations

Various demos highlighted Grok‑4’s real‑world prediction abilities, multimodal understanding, information integration, code and visualization generation, and voice synthesis.

World Series champion prediction on Polymarket, combining odds with FanGraphs data.

Identification of the "weirdest" employee avatars on X, showing subjective concept comprehension.

Timeline extraction from X posts for the "Humanity's Last Exam" scores.

HTML animation of two colliding black holes generated from a physics prompt.

Voice Mode and New Voices

Latency was halved, and five expressive voices were introduced, including the epic "Sal" and emotive British "Eve," which performed an improvised opera about diet soda.

SuperGrok Heavy Subscription

A premium tier called SuperGrok Heavy was announced at $300 per month, enabling Grok‑4 Heavy with parallel agents that collaboratively solve complex problems.

New Benchmarks

On the ARC‑AGI leaderboard, Grok‑4 surpassed o3 in version 1 and doubled it in version 2, though it still lags behind o3‑pro. In the Vending‑Bench commercial scenario, Grok‑4’s net value exceeded the second‑place model by more than twice.

Future Roadmap (45:32‑50:20)

August: Dedicated coding model.

September: Multimodal agents.

October: Video generation model.

The presentation concluded with a focus on AI safety, emphasizing the pursuit of truth, and ended with the classic line "So long, and thanks for all the fish."