grok-4.6

Overview. Grok 4.6 is xAI's closed-weight, reasoning-tuned flagship model, released in August 2026 with a training cutoff in February 2026. It accepts text and image input and produces text output, and it ships with mandatory reasoning that developers tune across four effort levels — low, medium, high, and xhigh — defaulting to high. Its 500,000-token context window makes it a candidate for developers who need to reason over large codebases or long documents in a single pass rather than a fast conversational model.

Key strengths. Grok 4.6 scores strongest on coding, reflected in an Artificial Analysis Coding Index of 76.8 and a curated code domain score of 0.77, with a general Intelligence Index of 60.9 and a math figure of 58.7. Because reasoning is mandatory and always on, the model consistently produces explicit reasoning traces rather than optional chain-of-thought, which supports multi-step problem solving over single-shot answers. Its combination of a 500,000-token context window and configurable reasoning depth suits workloads where thoroughness matters more than turnaround time.

When to use it. Reach for Grok 4.6 on long-running agentic coding or research tasks — large refactors, multi-file analysis, or research synthesis — where you can supply substantial context and want the model to reason through the problem at a chosen effort level. It's a good fit when the 500,000-token window lets you pass an entire codebase, document set, or research corpus in one request instead of chunking it. It also handles image input alongside text, so tasks that combine visual references with long-form reasoning are within scope.

Limitations. Grok 4.6 is closed-weight: xAI discloses no parameter count, architecture, or open-weights option, so it can't be self-hosted or inspected. It runs slower than average for its class and is best treated as a batch-latency model rather than one for tight, interactive turnaround, and it has no audio input or output. Its training data cutoff is February 2026, so it won't have knowledge of events after that point.

Sovereignty note. Grok 4.6 is created by xAI, a US-based organisation subject to CLOUD Act exposure, and is served as a proprietary, closed-weight model rather than a Sovereign, EU-resident one. Data residency and training-on-input policy for this model follow xAI's standard terms rather than EU-only guarantees.