Excel Bench | Entendre Research

Overview

Entendre Excel Bench evaluates whether frontier models can complete finance tasks using spreadsheets and other source material. Tasks test data retrieval, calculation, file preservation, and the quality of the final deliverable.

The benchmark contains 25 tasks: 21 require Excel workbooks, three require Word documents, and one requires a text answer. Agents receive task instructions and the available source files. Some tasks require external research. Each task specifies the required output and whether an existing workbook must be preserved.

The tasks span financial reporting, trading and risk, planning, asset analysis, and company research. They require models to use source evidence, calculate, structure outputs, preserve requested formatting, and validate results.

Tasks are authored and reviewed by finance workflow experts, so each task reflects realistic finance work and each evaluation check compares the final deliverable against clear analyst expectations.

Entendre Excel Bench measures execution quality through task-specific criteria for calculation accuracy, source use, output structure, formatting, and validation.

Model Scores

Model

Mean score

Score

Fable 5

76.0%

GPT-5.6 Terra

74.9%

GPT-5.6 Sol

74.7%

GPT 5.5

74.3%

GPT-5.6 Luna

71.8%

Sonnet 5

70.6%

Opus 4.8

70.2%

Gemini 3.5 Flash

68.3%

GLM 5.2

62.7%

DeepSeek V4 Pro

62.6%

Gemini 3.1 Pro

59.9%

Qwen 3.7 Plus

55.7%

Kimi K2.6

54.4%

0%25%50%75%100%

We measured three performance metrics:

Mean score: the average fraction of weighted grading criteria satisfied across all evaluated runs.
Best@3: the highest score from three attempts for each task, averaged across all 25 tasks.
Pass@3: the share of tasks where at least one of three attempts achieves a score of 90% or higher.

Methodology

This page reports a 13-model evaluation snapshot. Each model has three attempts on each of the 25 tasks, for 975 evaluated attempts. Agents use files and command-line tools to produce the requested deliverables. Results reflect each model's agent configuration, including its tool access and reasoning settings.

GPT-5.5 grades the outputs against task-specific reference material and weighted rubrics. Each criterion receives 1 for Passed, 0.5 for Partial, or 0 for Failed. The attempt score is the sum of each criterion's weight multiplied by its score, divided by the total weight.

Mean score includes all recorded attempts, including failed attempts. This snapshot contains all 975 evaluation records and no recorded evaluation errors. Best@3 averages the highest score for each task. Pass@3 is the share of tasks with at least one attempt scoring 90% or higher.

Scores measure performance on these 25 tasks under the evaluation setup. The judge evaluates both content and presentation, so a deduction can reflect formatting or incomplete evidence as well as an incorrect calculation. The capability analysis below covers five of the 13 leaderboard models.

Model Behavior

This capability analysis covers GPT 5.5, Sonnet 5, Opus 4.8, Gemini 3.5 Flash, and Gemini 3.1 Pro. It describes those five models rather than the full leaderboard.

Models can calculate, but finished workbook execution is still fragile.

Within this five-model comparison, GPT 5.5 leads workbook output, auditability checks, and data grounding. Sonnet 5 leads calculation and format fidelity. These category results describe specific capabilities and do not establish an overall winner across the 13-model leaderboard.

The lowest band is format fidelity. This is where small finance details matter: row placement, labels, number formats, and workbook presentation.

Stronger

Workbook output and calculation are the most dependable parts of the benchmark.

Weaker

Exact formatting, source discipline, and auditability checks are where the model behavior becomes less reliable.

Capability scores by model

GPT 5.5Sonnet 5Opus 4.8Gemini 3.5 FlashGemini 3.1 Pro

100%75%50%25%0

Workbook output

Preserve source workbooks and return a usable artifact.

Calculation

Compute values, totals, and spreadsheet-derived measures.

Auditability checks

Check outputs, reconcile evidence, and keep work auditable.

Data grounding

Use the correct source sheets, fields, and input records.

Format fidelity

Match labels, layout, number formats, and presentation details.

Example Task

One finance workbook task with example runs from five models. Inspect the task prompt, grading rubric, results, and agent trajectories.

Select model

This example uses strict reference matching. The agent receives the task instructions and source workbook. The reference solution and rubric are reserved for grading. The source workbook provides existing labels and formatting conventions, while the rubric also checks exact output row positions and number formats. A lower score can reflect differences from the reference format as well as incorrect calculations.

77.5%

Reward

gpt-5.5

Model

26

Steps

Prompt

Using Income-Qtrly as the base, please create the Income-Annual and Income-YTD schedules, formatted consistent with the other spreadsheets in the workbook.

Constraints

Modify the existing workbook in place. Use Income-Qtrly as the base, preserve all original sheets and workbook contents, add the requested Income-Annual and Income-YTD schedules, and return the full updated workbook rather than a results-only file.

Source files

Instruction fileWorkbook input

Rubric

Weight 45Data correctness

Income-Annual and Income-YTD must match the reference workbook displayed values. Annual values are sums of the four reported quarters per year; YTD 2008 values are running cumulative values.

Check every displayed line-item cell, with special emphasis on Annual Operating Revenues 2007=67920 and 2008=68345, Annual Net Income 2007=3892 and 2008=2832, Annual Total Operating Expenses 2007=53200 and 2008=61110, Annual Operating Income 2007=14720 and 2008=7235, YTD Operating Revenues 16620/33575/50885/68345, and YTD Net Income 2336/2588/4315/2832. EPS is judged on displayed two-decimal values where formatted.

Penalize every wrong displayed line-item cell; numerically correct but wrong-source or non-cumulative answers should lose heavily.

Weight 18Number-format consistency

The two new sheets must reproduce the workbook Accounting number-format family, rather than simplified currency or generic number formats. Dollar rows such as Operating Revenues and Net Income use the workbook dollar Accounting format.

Plain rows use the workbook plain-number Accounting format. EPS rows use the workbook EPS Accounting format.

Preserve the workbook quirk where Income Before Discontinued Operations and Loss on discontinued operations rows use the single-dash variant rather than normalizing all formats. Simplified currency, plain-number, or generic decimal formats should lose most of these points.

Weight 14Label fidelity

Labels must match the reference exactly, including deliberate typos and legacy terminology. Preserve "Selling, general & adminstrative expense", "Asset impairement and other charges, net", and "Minority interest".

Do not modernize to Noncontrolling interest. Drop "(Loss)" on the annual/YTD positive lines exactly as the reference does, including Operating Income, Net Income, Basic/Diluted Earnings per Share, and the Income Before discontinued/cumulative-effect lines.

The discontinued-ops unit name must read "Loss from operations of Telecom Andina".

Weight 13Layout fidelity

Row and column placement must match the reference. Income-Annual has revenues at row 6, EPS at rows 31 and 34, footnotes at rows 38-40, and dimensions 40x7.

Income-YTD has quarter labels on row 5, revenues at row 7, EPS at rows 32 and 35, footnotes at rows 38-39, and dimensions 39x12. A2 is a black-fill banner with white bold text.

Net Income values carry thin-top plus medium-bottom borders. YTD title uses an en dash in "Consolidated Statements of Income - Year to Date" equivalent to the reference text, year header 2008 uses centerContinuous over the period block with numeric 0 format, quarter labels keep irregular spacing ("1st Quarter", "2nd Quarter", "3rd Quarter", "4th Quarter"), and units notes sit right-aligned at G3/K3.

Weight 10Source discipline

Use Income-Qtrly reported data as the source, not Adjusted-Income sheets or any adjusted figures. Both target sheets must be present, exactly named Income-Annual and Income-YTD, and placed after Note/Income-Qtrly in the same order as the reference.

YTD cumulation must be true running cumulative logic; the discontinued-ops lines are a trap, with Loss from operations of Telecom Andina equal to -6 then -961 flat for Q2-Q4. Wrong-source adjusted revenue, missing/misnamed target sheets, deleted original sheets, or broad unrelated edits should be punished severely.

Previous benchmarks

May 21, 2026

TaxBench: US Federal Tax Return LLM Benchmark

A benchmark for evaluating how frontier AI systems perform on end-to-end US federal tax return preparation tasks.

Read benchmark