xAI released Grok 4.5, positioned as an Opus-class model that is faster, more token-efficient, and cheaper. It is worth separating what comes from the official announcement, what comes from independent evaluation, and what is specific to the launch partnership with Cursor.
Price and specifications
Grok 4.5 ships a 500k-token context window. The pricing: $2 per million input tokens, $6 per million output tokens, with cached input at $0.50 per million. That puts the model roughly 4x cheaper than Claude Opus 4.8 on output tokens, and over 8x cheaper than Claude Fable 5.
The model supports the full modern tool-use stack: native function calling, structured JSON output, web search, X search, and code execution, with "high" reasoning mode always on, no option to turn it off.
What independent evaluation shows
Artificial Analysis, which evaluates models independently of the vendor, ranked Grok 4.5 #4 among 168 models on its Intelligence Index, scoring 54. That is near the frontier, but behind Fable 5, GPT-5.5, and Opus 4.8 on raw quality according to that specific metric.
The distinction matters: "Opus class" in marketing and "#4 behind Opus" in independent evaluation are not the same claim. The model is competitive, not necessarily equivalent in pure quality.
Where the efficiency argument holds up
The most interesting metric from the launch is token efficiency per task, not absolute quality. On SWE-Bench Pro, xAI reports Grok 4.5 used 4.2 times fewer output tokens than Opus 4.8 to complete equivalent tasks: roughly 15,954 tokens versus 67,020. If that ratio holds in real workloads, the effective cost per completed task looks more favorable than the per-token price difference alone suggests, because fewer output tokens mean lower cost even at a comparable per-token price.
A caveat: that number comes from xAI itself, on a specific benchmark. Token efficiency varies by task type, and it is worth reproducing the measurement on your own workload before deciding based on it.
The Cursor launch
Cursor launched Grok 4.5 in partnership with xAI, describing it as their first model built for more than software engineering: long, complex tasks in data science, finance, and legal work, among other domains. Training used a mixture-of-experts architecture on trillions of tokens of real developer-agent interaction data collected within Cursor, with reinforcement learning on deliberately difficult problems, including distributed agent environments built to simulate complex tasks at scale.
An unusual bit of transparency: Cursor disclosed that its own codebase data was accidentally included in the training set, giving the model an unfair advantage on the company's internal benchmarks. Worth factoring in that bias when interpreting comparisons published by Cursor itself.
Pricing on Cursor: base model at $2/million input and $6/million output, fast variant at $4/million input and $18/million output. Available on desktop, web, iOS, CLI, and SDK.
What to take into a routing decision
For teams routing between models by cost, Grok 4.5 is a solid candidate for high-volume workloads where token efficiency per task matters more than the marginal quality gain at the top of the ranking. It is not the obvious choice when a task demands the maximum capability available, where models ahead on the Intelligence Index remain the reference. We should measure against our own context: not every vendor benchmark reproduces on your workload. How do you decide today between top-tier quality and cost per task when choosing a model?
Fonte: Introducing Grok 4.5
Top comments (0)