Can AI do the job?
AI models are tested on the same small units of professional knowledge and skill (knobits) as people. Pick an occupation, an activity or a single knobit to see which models can do it, and where people still lead.
Leaderboard: data modelling
Best result for data modelling: gemini-4-argon (89% test score).
| # | Model | Provider | Tier | Competence mastered (expected) | Knobits mastered | Test score | Cost per test run |
|---|---|---|---|---|---|---|---|
| 1 | gemini-4-argon | Flagship | 93% | 167/199 | 89% | $0.062 | |
| 2 | claude-fable-5-1 | Anthropic | Flagship | 92% | 148/199 | 88% | $0.85 |
| 3 | claude-sonnet-5-5 | Anthropic | Mid tier | 91% | 145/199 | 88% | $0.23 |
| 4 | gpt-6-astra | OpenAI | Flagship | 90% | 152/199 | 87% | $0.45 |
| 5 | gpt-6.1-sol | OpenAI | Mid tier | 88% | 144/199 | 85% | $0.077 |
| 6 | gemini-3.8-flash | Mid tier | 88% | 143/199 | 80% | $0.045 | |
| 7 | claude-haiku-5-5 | Anthropic | Low cost | 86% | 148/199 | 79% | $0.018 |
| 8 | gpt-6-luna | OpenAI | Low cost | 82% | 110/199 | 75% | $0.005 |
| 9 | gemini-3.1-pro-preview | Flagship | 80% | 109/199 | 76% | $0.20 | |
| 10 | gemini-3.5-flash-lite | Low cost | 65% | 62/199 | 54% | $0.008 |
AI model statistics per activity
Expected share of each activity's knobits a model masters. Until at least 5 operators have shared results for a model and an activity, the figure is from Skilirium's benchmark (3 runs per model, the same conditions for every model). From then on it is from operators' own agents in their work settings (a model with their own instructions, tools and data, self-reported), marked †. Each figure has one source; the two are never mixed.
No model has shared results from 5 or more operators yet, so every figure is from the benchmark. Operator results exported 2026-10-29. Download the figures (JSON)
| Model | Capture data requirements | Define business terms | Document data rules as SBVR fact types and rule statements | Build the conceptual model and taxonomies | Derive the logical data model | Design the physical data model | Derive data quality rules | Generate and review DDL |
|---|---|---|---|---|---|---|---|---|
| gemini-4-argon | 88% | 88% | 97% | 100% | 95% | 98% | 95% | 82% |
| gemini-3.1-pro-preview | 52% | 81% | 86% | 100% | 94% | 84% | 86% | 61% |
| gpt-6-astra | 78% | 94% | 94% | 87% | 91% | 97% | 98% | 81% |
| claude-fable-5-1 | 78% | 99% | 98% | 100% | 94% | 98% | 95% | 73% |
| gemini-3.8-flash | 67% | 94% | 93% | 100% | 94% | 92% | 94% | 66% |
| gpt-6.1-sol | 73% | 94% | 94% | 88% | 89% | 92% | 98% | 78% |
| claude-sonnet-5-5 | 75% | 98% | 100% | 94% | 95% | 94% | 92% | 76% |
| gemini-3.5-flash-lite | 53% | 68% | 53% | 92% | 66% | 68% | 66% | 56% |
| gpt-6-luna | 59% | 73% | 87% | 100% | 92% | 90% | 89% | 64% |
| claude-haiku-5-5 | 70% | 82% | 93% | 100% | 94% | 92% | 95% | 64% |
Browse
- Data modeller · 130 of 307 knobits mastered by most models
- Assessment specialist · not benchmarked yet
- All knobits
- All AI models tested
How the models are tested
- 10 test tasks, built around one realistic case (a loan application), each assessing one or two activities. Each model answers each task 3 times, at its default settings, through its provider's API.
- Every answer is graded against the task's rubric by claude-opus-5-5, plus automatic checks. The judge model is not among the models tested.
- Each rubric criterion names the knobits it needs. A criterion counts as solved at 75% of its points.
- From the solved and failed criteria, knowledge space theory estimates each model's probability of mastering every knobit, as it would for a person. A knobit is mastered at a probability of at least 90%.
- Costs are the mean API cost of one test run at list prices; they compare models, not budgets.
- Last test run: 2026-10-14.
Caveats
- Skilirium is a pilot. The tests were drafted with the help of AI and are being reviewed by data modelling experts: read the results as an indication, not a certificate.
- A judge model may favour answers from its own model family; keep that in mind when comparing Claude models with the others.
- Repeated answers from one model are not independent, and the judge scores whole criteria, not their parts.
- Mastering knobits in a test is not doing the job: ownership, judgement with stakeholders and accountability stay with people.