Grok 4.5

Grok 4.5

Quick answer: Grok 4.5 is xAI's flagship model released 8 July 2026, designed for code, agentic tool calling, and general reasoning. It ships with a 500K-token context window, configurable reasoning effort, and a low hallucination rate. Pricing is $2.00 input / $6.00 output per 1M tokens. Knowledge cutoff is February 1, 2026.

At a Glance

Where Grok 4.5 leads

• Top performance on long-horizon terminal tasks — #1 on LHTB with 0.505 mean reward and 13/46 tasks solved.
• Strong cost-to-performance ratio at $2/$6 per 1M tokens, delivering the best LHTB score at ~$11 per 90-minute task.
• Configurable reasoning effort (non-reasoning and reasoning modes) for flexible deployment.

Where it lags

• 500K context window, smaller than the 1M windows of Grok 4.20, Claude Opus 4.8, and DeepSeek V4 Pro.
• Closed, API-only model with no open weights.

Best for: Long-horizon agentic coding, tool-heavy workflows, and cost-efficient frontier performance.

What Grok 4.5 Is

Grok 4.5 is xAI's current production flagship, released to replace Grok 4 as the recommended model for code, chat, and agentic tasks. According to xAI, it leads the industry in coding capability and agentic tool calling, while also posting the lowest hallucination rate among frontier models. Like its predecessors, it supports function calling and structured outputs, and can be combined with xAI's Web Search and X Search tools for real-time grounding.

The model ships with configurable reasoning: developers can set the reasoning effort to a non-reasoning mode for latency-sensitive tasks or enable the full reasoning path for problems that benefit from deliberate chain-of-thought. Its 500K context window is well suited for large codebase traversal, long documents, and multi-turn agentic sessions, though it falls short of the 1M-token windows offered by some competing models.

On the LHTB benchmark — which measures agent performance on 46 hard, multi-hour terminal tasks — Grok 4.5 scored 0.505 mean reward, placing it first overall. It solved 13 of 46 tasks outright (reward ≥ 0.95) and did so at roughly $11 per task, making it the best-performing model at the lowest cost among the frontier options evaluated.

Specifications

Related Benchmarks

• Long-Horizon Terminal-Bench (LHTB) — 0.505 mean reward, rank 1 (13/46 solved)
• TerminalBench 2.1 — part of the terminal agent benchmark landscape