BizBench

Quantitative finance and accounting reasoning benchmark using program synthesis over SEC tables.26 Oct 2026

One row per model and compatible score measure. Mean = equal-weight arithmetic average of imported reports. Distinct configurations reported for the same model in a source table remain separate. Reports can overlap; the range is not a confidence interval. Unbundle to inspect each observation.
		Evidence	What the result shows
GPT-4* / - →OpenAI	84.6 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	Leads this displayed comparison. Source notes Source cell: 84.6
GPT-3.5* / - →OpenAI	72.5 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#2 of 16 in this displayed comparison; 12.1 points below the leader. Source notes Source cell: 72.5
GPT-3* / 175B →OpenAI	72.3 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#3 of 16 in this displayed comparison; 12.3 points below the leader. Source notes Source cell: 72.3
Mixtral / 8x7B →Mistral	69.9 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#4 of 16 in this displayed comparison; 14.7 points below the leader. Source notes Source cell: 69.9
Llama 2 / 70B →Meta	68.4 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#5 of 16 in this displayed comparison; 16.2 points below the leader. Source notes Source cell: 68.4
CodeLlama / 34B →Meta	62.5 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#6 of 16 in this displayed comparison; 22.1 points below the leader. Source notes Source cell: 62.5
Mistral / 7B →Mistral	61.8 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#7 of 16 in this displayed comparison; 22.8 points below the leader. Source notes Source cell: 61.8
CodeLlama / 13B →Meta	56.8 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#8 of 16 in this displayed comparison; 27.8 points below the leader. Source notes Source cell: 56.8
Llama 2 / 13B →Meta	54.7 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#9 of 16 in this displayed comparison; 29.9 points below the leader. Source notes Source cell: 54.7
MPT / 30B →	53.5 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#10 of 16 in this displayed comparison; 31.1 points below the leader. Source notes Source cell: 53.5
CodeLlama / 7B →Meta	52.1 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#11 of 16 in this displayed comparison; 32.5 points below the leader. Source notes Source cell: 52.1
Llama 2 / 7B →Meta	48.2 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#12 of 16 in this displayed comparison; 36.4 points below the leader. Source notes Source cell: 48.2
StarCoder / 16B →	47.5 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#13 of 16 in this displayed comparison; 37.1 points below the leader. Source notes Source cell: 47.5
Falcon / 40B →	41.3 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#14 of 16 in this displayed comparison; 43.3 points below the leader. Source notes Source cell: 41.3
MPT / 7B →	37.1 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#15 of 16 in this displayed comparison; 47.5 points below the leader. Source notes Source cell: 37.1
Falcon / 7B →	28 points 1 report	BizBench · Published task average[1]BizBench · zero-shot and few-shot evaluation	#16 of 16 in this displayed comparison; 56.6 points below the leader. Source notes Source cell: 28.0

Sources

BizBench: A Quantitative Reasoning Benchmark for Business and Finance — published result table ↗
Rik Koncel-Kedziorski, Michael Krumdick, Viet Lai, Varshini Reddy, Charles Lovering, Chris Tanner · 2023
Dates & evaluation details
BizBench · Published task average · zero-shot and few-shot evaluation · 2026-10-26
Published model comparison, 2023. BizBench · zero-shot and few-shot evaluation. The source average combines the named domain-knowledge, extraction and program-synthesis tasks with their published shot settings. This is not live financial performance. Table 3; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

TRACE THE SCORE · Per-model notes use published breakdowns where available.

What it measures

Quantitative finance and accounting reasoning benchmark using program synthesis over SEC tables.

Reading the measurement

Check the source documents, dates, accounting conventions and calculation chain. A financial answer needs evidence and consistent units. Benchmark success does not establish live investment performance or suitability for an individual decision.

Published evaluation

The source average combines the named domain-knowledge, extraction and program-synthesis tasks with their published shot settings. This is not live financial performance.

Where this fits in an evaluation suite

Synthetic research rewards do not establish live investment performance. Preserve transaction costs, held-out regimes, constraints, and the exact policy behind a reported result.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

The source average combines the named domain-knowledge, extraction and program-synthesis tasks with their published shot settings. This is not live financial performance.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.

PRIMARY EVIDENCE & FURTHER READING

BizBench: A Quantitative Reasoning Benchmark for Business and Finance — published result table ↗
Rik Koncel-Kedziorski, Michael Krumdick, Viet Lai, Varshini Reddy, Charles Lovering, Chris Tanner · 2023

Reviewed 2026-10-26. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.

Continue exploring

All financial reasoning benchmarks →

Sources

What it measures

Reading the measurement

Published evaluation

Where this fits in an evaluation suite

A reproducible comparison

Continue exploring

τ³-Banking

BankerToolBench

OfficeQA Pro

Vals Finance Agent v2