Independent Tests Rank Grok 4.5 Fourth Among Top AI Models

Research firm Artificial Analysis has published independent benchmark results for Grok 4.5, SpaceXAI's model released on July 8, 2026. The model scored 54 points on the Intelligence Index, placing it fourth behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8, but it beat rivals on agentic tasks and turned out to be many times cheaper per completed task.

Fewer Points, More Efficiency

The 54-point Intelligence Index score marks a 16-point improvement for Grok 4.5 over the previous version, Grok 4.3, and puts SpaceXAI's model at the edge of the world's top systems, though still behind the three leaders from Anthropic and OpenAI. Grok 4.5 does outpace every open-weight model as well as Google's Gemini family, which is itself a meaningful shift in the market's balance of power.

The model looks far more interesting on token-efficiency tests. On SWE-Bench Pro, Grok 4.5 solves tasks using an average of 15,954 output tokens, while Claude Opus 4.8 in max mode needs as many as 67,020 tokens for the same task, more than four times as many. At a speed of 85.6 tokens per second, against a market average of around 73, Grok 4.5 ranks among the fastest models currently offered commercially.

Agentic Tasks and Banking

The model's strongest suit turned out to be agentic tasks, scenarios that require carrying out multi-step actions independently without constant human oversight. On the AutomationBench-AA benchmark, Grok 4.5 posted the best result of any model tested, ahead of both Claude Fable 5 and Claude Opus 4.8. Likewise on τ³-Banking, which simulates customer service in the banking sector, SpaceXAI's model logged the highest score of all the systems measured.

Artificial Analysis notes that Grok 4.5 sits on the cost-effectiveness frontier across all three agentic tests at once, meaning no other model offers a better quality-to-price ratio in this category of tasks. That matters for companies deploying AI agents for customer service or process automation, where the cost of a single completed task counts for more than a raw general-knowledge score.

Price as the Main Argument

At $2 per million input tokens and $6 per million output tokens, Grok 4.5 costs $0.31 per task on the Intelligence Index, many times less than pricier rival models with comparable capabilities. On coding tests, the cost per task on the Coding Agent Index is $2.59, and the model itself scores 76 points there, on par with GPT-5.5 in high-performance mode and only slightly below Claude Fable 5.

That combination of high agentic performance and low cost changes the economics for companies running AI agents at scale. Since models are billed per token rather than per query, a fourfold difference in token consumption per task translates directly into infrastructure bills running into millions of dollars a month at high volume.

Grok 4.5 stands out on the Pareto frontier across all three agentic evaluations - Artificial Analysis

A Shadow of Controversy

The launch was also accompanied by discussion unrelated directly to the model's capabilities. The loudest thread on Hacker News on release day concerned not performance but trust, with users pointing to earlier allegations that Elon Musk had interfered with how Grok models answer political questions. It is another episode in a string of controversies over content moderation and bias in this model family that have followed the brand since its founding.

For Polish companies and developers weighing which language model provider to choose for agentic tasks, the Artificial Analysis results show that ranking position on general intelligence alone doesn't tell the whole story. In concrete business applications like customer-service automation or large-scale coding tasks, the cost of completing a task may matter more than a raw knowledge-test score.

Sources: Artificial Analysis (artificialanalysis.ai), MarkTechPost (marktechpost.com), Tech Times (techtimes.com)