The Practical Work Benchmark

Cover of The Practical Work Benchmark PDF · 15 pages ↓

This report compares seven leading AI models on the practical work knowledge workers do every day. It is a structured review of public evidence: 82 source records screened and turned into 168 traceable claims, with evidence frozen August 9, 2026. Its main finding is that all seven models are strong, and their costs differ widely.

What’s inside

  • Six jobs. Automating repetitive reporting and workflows, analyzing data for decisions, drafting professional communication, writing documentation, measuring performance and ROI, and integrating systems and troubleshooting problems.
  • The models. GPT-5.6 Sol, Claude Opus 5, Gemini 3.6 Flash, Grok 4.5, Kimi K3, GLM-5.2 and DeepSeek-V4-Pro. Three of them (Kimi K3, GLM-5.2, DeepSeek-V4-Pro) have open weights.
  • A decision matrix. Each model is rated from “Just OK” to “Incredible” on each job, with a marker for how much evidence supports each rating.
  • Capability against cost. On DocBench, all seven fall within an 85 to 100 relative rating band. List prices range from $0.435 input and $0.87 output per million tokens (DeepSeek-V4-Pro) to $5 and $30 (GPT-5.6 Sol).
  • Intelligence per dollar. A rough measure of output per dollar, adjusted for quality. DeepSeek-V4-Pro leads at 975.9K, then GLM-5.2 at 204.2K. GPT-5.6 Sol is lowest at 33.3K.
  • Model profiles. For each model: what it is best suited for, watchouts, and quotes from practitioner tests. For example, Sol leads on finished editable documents, while Claude Opus 5 is strongest at analysis and prose but can be verbose.

Key takeaway

Any of these models would have looked exceptional a year ago. The buying decision now comes down to price, measured as intelligence output per dollar spent.

Bring your agent. Bring your team.