What happens when you judge AI models not on SAT questions or coding puzzles, but on whether they can run a business?
A team put 12 language models through a year of simulated startup operations - setting budgets, hiring staff, making product decisions, and managing cash flow under competitive pressure. Claude Opus 4.6 ranked first overall. But GLM-5, a model from Zhipu AI, a Chinese lab backed by Tencent and others, finished close behind - at approximately 11 times lower API cost.
What the Test Actually Measured
Unlike standard benchmarks that test factual recall or code completion, this simulation required models to make interdependent business decisions across multiple time steps. A bad hiring decision in month 3 would compound into worse outcomes by month 8. Models were scored on business results - revenue growth, cost management, competitive positioning - not on whether they picked the right answer to a discrete question.
Real business work involves chains of decisions where early mistakes compound. Most AI benchmarks don't capture this because they treat each question in isolation. This one did, and the ranking it produced looks different from the standard leaderboards.
The GLM-5 Cost Gap
At 11x lower API cost than Claude Opus 4.6, GLM-5's near-parity performance matters for anyone building AI-powered business automation. The benchmark doesn't mean GLM-5 outperforms Claude at everything - it still ranked below Claude Opus 4.6 overall. But for business reasoning specifically, the gap appears small relative to the price difference.
GLM-5 lacks the broad tool integrations and Western platform support that Claude and GPT-4o have. Running it requires API access to Zhipu's service or self-hosting, which adds operational complexity. But cost-sensitive applications - customer support automation, financial planning tools, business analytics pipelines - now have a serious alternative worth evaluating.
One Data Point
One benchmark is one data point. This simulation tested a specific type of business reasoning across a narrow scenario. It says nothing about writing quality, code generation, or the many other tasks that real AI deployments handle. "Nearly matched" is also doing some work in the headline - the actual score gap and the methodology would need independent replication before drawing strong conclusions.
But the study reinforces something the industry is learning fast: the gap between top-tier and second-tier models is closing on cost, and the models dominating standardized test leaderboards don't always win when the test looks more like actual work.