100-task capstone
Accounting Capstone
Hybrid scoring · 17 published model runs
View benchmark →Independent financial AI evaluation
Which AI models can actually perform professional financial work?
Kaelum Bench evaluates accounting, reporting, and corporate finance with expert-designed tasks and signed evidence attached to every published model run.
v0.2 · Public CalibrationUpdated August 14, 2026Signed public release
Top models this release
v0.2 · signed release
- Kimi K2.688.9
- Kimi K388.9
- Claude Opus 588.5
- GPT-5.6 Terra88.4
- Claude Sonnet 5 (Max)86.1
Mean of nine benchmark scores · All 15 models →
Current release metrics
9 Published benchmarks
of 10 defined
580 Tasks represented
across both professional directions
164 Verifiable run receipts
132 deterministic · 32 hybrid
Latest benchmark release
The current calibration, clearly published.
v0.2 publishes the focused accounting and corporate-finance benchmarks alongside the first integrated 100-task capstone.
v0.2 · Signed public release
Accounting and corporate finance, evaluated as professional work.
Explore focused 60-task benchmarks and the published Accounting Capstone. Every observed score remains connected to methodology, configuration, cost, token use, and exact run evidence.
Open all benchmarks →Next 100-task capstone
Corporate Finance Capstone
Benchmark definition prepared; calibration pending
Preview scope →Choose your depth
Start with the question you have.
Explore benchmarks
Scan every published benchmark, focused suite, and 100-task capstone in one consistent overview.
Explore benchmarks →Compare models
Select models and compare observed benchmark performance and recorded resources across the release.
Compare models →Understand the method
See what is tested, how points are awarded, and which limits matter when reading a score.
Understand the method →Transparent by design
Evidence before comparison.
Task and scoring evidence
Definitions, task counts, scoring modes, and tested scope remain visible.
Execution and cost evidence
Provider route, runtime, tokens, cost, and validity stay attached to each run.
Signed publication
Versioned data and authenticated manifests preserve the exact public release.