xAI опубликовала model card для Grok 4.5: бенчмарки и оценки безопасности

xAI published the official model card for Grok 4.5 in July 2026 — a document that gathers the model's results on capability benchmarks and the outcomes of its safety evaluations. The team unveiled the Grok 4.5 model itself on July 8, 2026, and the card adds detailed technical documentation to the release.

What is a model card

A model card is a technical passport for a neural network: a short official document that a model's developer publishes to show what it can do and where its limits lie. For Grok 4.5, xAI's card brings together two blocks of data — capability benchmark measurements and safety evaluation results.

• The model is Grok 4.5, the flagship version of the Grok family
• Grok 4.5 was unveiled on July 8, 2026
• The document is an official model card from the development team
• The content covers capability benchmarks and safety evaluations
• The format is a public document released alongside the model's launch

What the card documents

The Grok 4.5 card records two types of results: measurements on standardized capability benchmarks and the outcomes of safety checks. A benchmark is a set of tasks with known correct answers, used to measure how well a model reasons, writes code, and works with text; these figures are used to compare the new version with previous versions and with competitors. The second block, safety evaluations, shows how the model behaves in risky scenarios and what limitations the developers built into it.

"We are publishing the official model card for Grok 4.5.

It documents the model's capability benchmarks and safety evaluations," the team's blog post accompanying the release states.

Why publish safety evaluations

Publishing safety evaluations alongside benchmarks is a way to show not just the model's strengths but also its weak points before people start using it. Without this block, users see only the strong sides and don't know where the model might make mistakes or behave unpredictably. Releasing model cards has become an industry standard: major AI labs do this so developers and regulators can assess risks in advance. For Grok 4.5, both blocks — capabilities and safety — are gathered in a single document published in July 2026.

What this means

The model card shifts the Grok 4.5 announcement from a marketing plane to a verifiable one: instead of general phrases about the "most powerful model," there are now concrete capability measurements and risk assessments. For teams building products on Grok, this is a reference document — it's used to decide whether the model fits a specific task and how much trust to place in its answers.

Frequently asked questions

What does the Grok 4.5 model card show?

The Grok 4.5 model card shows two blocks of data: the model's results on capability benchmarks and the outcomes of safety evaluations. It is an official document published alongside the release in July 2026.

When was Grok 4.5 unveiled?

Grok 4.5 was unveiled on July 8, 2026. The official model card with benchmarks and safety evaluations was released as part of that release.