Accounting · Agentic environments · August 2026
Every model in the field fails at least one task outright.
Per model
The distance between passing the check and doing the work.
Steady credit is not a pass. Nothing here separates statistically, so read it as directional.
ModelRuns cleared, of 15Strict
kimi-k38 / 15rubric 0.78
grok-4.57 / 15rubric 0.81
gemini-3.6-flash6 / 15rubric 0.77
gpt-5.6-sol6 / 15rubric 0.58
gpt-5.6-luna3 / 15rubric 0.42
claude-opus-53 / 15rubric 0.68
gemini-3.1-pro3 / 15rubric 0.52
Engine verifierStrict passThe gap
Where the roadmap heads
Investment banking is the same method with more at stake.
Buyout and merger modelling, comps, diligence. A hallucinated number in a live deal costs more than one in a monthly close.