# JobBench ![Category](https://img.shields.io/badge/Category-Workforce%20Agent-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20Success-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) ![Created](https://img.shields.io/badge/Created-2026-555) > **Quick answer:** JobBench evaluates AI agents on tasks drawn directly from real job roles across a range of occupations, testing whether a model can perform the actual day-to-day work associated with specific professions. Kimi K3 scores 54.3% as of July 2026. ## At a Glance **What it tests:** An agent's ability to complete authentic job-role tasks — the kind of work assigned to professionals in specific occupations — rather than abstract or synthetic test items. **Why it matters:** As AI agents are increasingly evaluated for workforce automation potential, JobBench offers a task-grounded way to measure real occupational capability rather than proxy metrics. **Known limitations:** As an emerging benchmark, the specific occupations and task sourcing methodology are not yet independently published outside its citation by Moonshot AI. ## What JobBench Measures JobBench evaluates an agent's ability to perform tasks sourced from real job roles across a range of occupations, testing occupational competence directly rather than through proxy academic tasks. This complements broader economically-valuable-task benchmarks (like Agents' Last Exam) by focusing specifically on job-role-grounded task completion. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Agent / workforce automation | | Metric | % task success | | Saturation | Low | | Created by | Not yet independently documented | ## How JobBench Is Scored Agents attempt tasks drawn from real job roles, with outcomes verified against expected deliverables, producing an aggregate % task success rate. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [Kimi K3](/models/moonshot-ai/kimi-k3) | 54.3% | [Kimi K3 technical report](https://www.kimi.com/blog/kimi-k3) | 2026-07 | *Score sourced from Moonshot AI's Kimi K3 announcement, July 2026.* ## JobBench on Benchgen > No Benchgen results yet — [be the first to run JobBench](/benchmarks/li-et-al/jobbench). ## JobBench vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **JobBench** | Real job-role task completion | Low | | [Agents' Last Exam](/benchmarks/berkeley/agents-last-exam) | Economically valuable long-horizon agent tasks | Low | | [APEX-Agents](/benchmarks/mercor/apex-agents) | Professional-grade knowledge-work agent tasks | Low | | [AutomationBench](/benchmarks/zapier/automationbench) | Business workflow automation | Low | ## Run JobBench on Your Model Benchgen lets you run JobBench against your own agent, tracking real job-role task success rates over time to assess workforce-automation readiness. ## Frequently Asked Questions

What is JobBench?

JobBench is a benchmark testing AI agents on tasks drawn from real job roles across a range of occupations.

What does a good score look like on JobBench?

Kimi K3 reports 54.3% as of July 2026, a solid frontier-level result reflecting genuine occupational task competence.

Who created JobBench?

JobBench's originating team is not yet independently documented outside of its citation in Kimi K3's July 2026 technical report.

Rank	Model	Score
1	mimo-v2-6-pro-rl	62
2	muse-spark-open-816b-a42b	61.8
3	hy4-preview	61.7
4	mimo-v2-6-flash-rl	61.2
5	qwen3-8-flash-next	55.7
6	kimi-k3	54.3
7	nex-n2-5-max	53.6
8	qwen3-8-max	53.4
9	nanbeige4-3-33b	44.8
10	nex-n2-5-pro	41.4
11	qwen3-8-27b	33.4

JobBench — Results

JobBench — Evaluations