Grok 4.6 made large gains on AA-Briefcase, our agentic knowledge work benchmark

Artificial Analysis's AA-Briefcase result for xAI's Grok 4.6 is a tie on quality and a large gap on cost (post). AA-Briefcase is the evaluator's benchmark of long-horizon agentic knowledge work, and its test set is private so that models cannot have seen the tasks in training (how the benchmark works). Artificial Analysis says Grok 4.6 is "neck and neck" with Anthropic's Claude Fable 5, with overlapping confidence intervals. That means the measured difference between them is within the benchmark's margin of uncertainty, so the result does not show which model is better.

Cost per task is where they separate. Grok 4.6 cost $4.42 per task, against $22.30 for Claude Fable 5, $17.79 for Claude Opus 5 and $6.73 for Kimi K3. Claude Fable 5 therefore cost about five times as much for a result this benchmark cannot distinguish from Grok's. These are the costs of running this benchmark's tasks, not general prices for knowledge work.

The post's other claim, that Grok 4.6 made "large gains," is not quantified in its text, and the post does not say what the gains are measured against. A later Artificial Analysis post on Grok 4.7 ranks that model behind only Anthropic's models on the benchmark.