Grok 4.5 hits #1 on Long-Horizon Terminal-Bench, lands in Perplexity same day

Grok 4.5, released July 8 and now powering Grok Build, is xAI's flagship coding and agentic model. Sources describe it as a mixture-of-experts system (Perplexity's Source B cites 1.5T parameters) trained on real Cursor agent-interaction data, with a 500,000-token context window and pricing of roughly $2 per million input tokens and $6 per million output. It scored 29.0% on the SWE Marathon benchmark, competitive with Opus-class models but tuned for cost and speed.

Elon Musk touted a fresh milestone this week — Grok 4.5 reaching the #1 spot on Long-Horizon Terminal-Bench — which reframes the model as a genuine agentic contender rather than a chat-first product. The commercial validation came fast: Perplexity CEO Aravind Srinivas said the model was wired into Perplexity Computer 'within a few hours' because it topped their internal evals, was the most cost-effective option, and offered zero-data-retention (ZDR) that enterprise customers demand.

Competitively, Grok 4.5 slots into the week's price war alongside GPT-5.6 Luna and Claude Sonnet 5, matching Opus 4.7-class speed at a fraction of the price. Cursor developers and X engineering teams call it a 'credible third option' for coding, though community testers note a roughly 2.8-point benchmark gap versus GPT-5.6 Sol. The launch also followed xAI's reported $250B merger with SpaceX and a floated 'SpaceXAI' rebrand — a corporate backdrop that adds capital but also raises questions about focus.