Elon Musk’s AI company, xAI, has launched Grok 4.7, the latest model in the Grok lineup. The model features 2.1 trillion parameters, a 40% increase over Grok 4.6’s 1.5 trillion, and was trained on supplementary data from SpaceX, including Starlink satellite telemetry, manufacturing records, and engineering issue logs. This marks the fifth model in the Grok 4 lineup, following at least five delays since late July.

Benchmarks: Still trailing Fable and Astra

According to benchmarks published by xAI, Grok 4.7 delivers competitive results but does not lead the pack.

  • GDPval (professional office work): Grok 4.7 scored 1,695 Elo, trailing Claude Fable 5.1 (1,735).
  • AA Briefcase (multi-hour office tasks): 1,657, trailing Fable 5.1 (1,678).
  • CursorBench 4.0 (coding): 46.3%, behind Fable 5.1 (51.8%) but ahead of GPT-5.6 Sol (41.7%).
  • EEBench (electrical engineering): Grok 4.7 scored 64.0%, well ahead of Fable 5.1 (56.4%) and GPT-5.6 Sol (39.4%).
  • Harvey Legal Agent Benchmark (legal): 19.6%, beating Fable 5.1 (6.7%). These are areas where SpaceX data may have made a difference.

Musk described Grok 4.7 as a “potent mix of intelligence, speed, and low cost.” But Musk himself had tempered expectations earlier, writing that Grok 4.7 would be “roughly on par” with Claude Opus 5.0, not Opus 5.1, and that multimodal performance still needed improvement.

Far cheaper than comparable US models

The real strength of Grok 4.7 is pricing. $2 per million input tokens and $6 per million output tokens, 5 times cheaper than Fable and 3 times cheaper than GPT-5.6 Sol for output.

This aligns with xAI’s consistent strategy: continually undercutting Anthropic and OpenAI despite lower capability, wagering that “good enough, cheap, and ubiquitous” beats “best but more expensive” for the vast majority of everyday users. Grok is currently available across X, a standalone app, the Tesla dashboard, Cursor, Grok Build, and the API.

New safeguards and SpaceX data

Grok 4.7 was built with an entirely new safeguard system, making it the strongest model xAI has tested in refusal rates and jailbreak resistance. On the LatchBio biosafety benchmark, the model scored 62.4%. On HackerBench v0.3 (a benchmark for hazardous cyber tasks), only 3.3% of risky prompts were permitted, while rarely blocking legitimate security work.

Regarding SpaceX data: xAI incorporated Starlink satellite telemetry, manufacturing records, and engineering issue logs into the training process. The goal is for the model to reason better about hardware and physical systems than models trained solely on internet text—an edge no other AI lab possesses.

Roadmap: Grok 4.8, 4.9, and 5

Musk outlined the next three models: Grok 4.8 as a significant step forward, Grok 4.9 at the “Astra/Fable level,” and Grok 5 potentially leading the frontier. These models do not yet have release dates. Musk wrote: “We shall see.”

For users, Grok 4.7 is a “good enough at a low price” model, not leading the benchmarks but ubiquitous wherever Grok users are. The practical question is not whether Grok 4.7 beats Fable, but whether being 5 times cheaper is enough to make up for the capability gap for daily tasks.