xAI: Grok 4.5 — new flagship for coding and agentic work at $2 per million input tokens

In its July changelog, xAI released Grok 4.5, which the company positions as its “flagship for coding and agentic work.” xAI’s release notes do not specify an exact release date — the entry is reported in the July changelog, and Perplexity’s changelog with model support confirms this is a fresh release.

Price as the primary weapon

Grok 4.5 costs $2 per million input and $6 per million output tokens. For comparison, comparable frontier coding models typically range from $3 to $15 for input and $15 to $75 for output per million tokens — xAI thus undercuts the competition by several times. For agentic tasks, where a single task routinely consumes tens of thousands of tokens, the difference directly flows through to product economics; a recent arXiv paper on the “Harness Effect” measures the average cost of an agentic task at 12 to 21 cents, precisely on this class of models.

How does configurable reasoning work?

The model supports three reasoning effort levels — low, medium, and high (default high). Reasoning effort determines how many tokens the model spends on its internal chain of thought before responding: higher effort means more accurate results on complex problems, but higher cost and latency. The same lever appeared in the GPT-5.6 family released the same week, confirming that configurable reasoning is becoming an industry standard.

Distribution through Perplexity

Grok 4.5 is immediately available in Perplexity’s Agent API as xai/grok-4.5 — a channel that gives xAI reach to developers outside its own platform. In a week when OpenAI launches GPT-5.6 and Meta releases Muse Spark 1.1, xAI re-enters the race with a pricing strike on the fastest-growing segment: agents that write code.