grok-4.5

Overview. Grok 4.5 is xAI's proprietary reasoning model, released on 8 July 2026. It accepts text and image input and returns text, over a 500,000-token context window. Inference runs on xAI's hosted API only. Developers reach for it when they want a reasoning model whose thinking trace is always on and whose effort level they can dial, rather than a fast non-reasoning chat endpoint.

Key strengths. Reasoning is mandatory on Grok 4.5 and configurable across low, medium, high and xhigh effort levels, with high as the default, so every response includes a chain of thought rather than hiding it. The 500,000-token context window suits long documents and large codebases, and the model supports function calling and structured outputs. The model's family record carries an arena Elo of 1470, placing it in a competitive tier, and its strongest domain signal is code.

When to use it. Use Grok 4.5 for reasoning-heavy work where latency is not the binding constraint: multi-step analysis, long-document review, and code tasks that span a large repository. Its measured domain score is highest in code, so it fits code-centric loads ahead of writing-heavy or conversational ones. Because reasoning is always on and the context window is large, it belongs in batch or asynchronous pipelines rather than realtime user-facing turns.

Limitations. The model is closed-weight, and xAI publishes no parameter count, architecture class or mixture-of-experts details, so there is no self-hosting path and no way to reason about its internals from outside. Training data ends on 1 February 2026. Input is limited to text and image, with no audio input or output, and the always-on reasoning trace means every call spends tokens on chain of thought before producing an answer.

Sovereignty note. Grok 4.5 is built by xAI, a United States company, and is offered only as a hosted closed-weight API, so workloads sent to it fall under US jurisdiction including CLOUD Act exposure. The research findings record no vendor residency guarantee or input-confidentiality commitment, so this model is not a fit for workloads flagged sensitive or Sovereign.