The benchmark for real work

CorpBench

Most benchmarks score answers. CorpBench scores the work itself: the final state, the actions taken, and whether the agent completed or escalated correctly.

16k
Tool calls
100
Workflows
5
Departments
18
Models
GitHub (opens in a new tab)Hugging Face (opens in a new tab)

01Key Findings

The results at a glance

  • The highest-scoring model is also the cheapest API option.

    GPT-6 Luna leads with a CorpBench Score of 94 and passes 98% of tested workflows at $0.00114 per task. Claude Haiku 5.5 passes 94% at around seven times the cost. Higher spending does not guarantee better work.

  • Speed and partial credit can hide unfinished work.

    Gemini 3.5 Flash-Lite averages 4.5 seconds per task, compared with GPT-6 Luna’s 16.2 seconds. Its average outcome score is nearly 98 out of 100, yet it passes only 83% of workflows. Check whether the whole job gets done before choosing on speed alone.

  • Local models vary widely in what they can finish.

    Gemma 4 E4B leads the tested local models with a CorpBench Score of 84 and a 77% pass rate. The other local models pass between 20% and 38%. Local runs have no API charges, but hardware and electricity costs still matter.

02Overall results

Which models perform best overall

Models are ranked on the quality of their work, how often they pass, and how efficiently they run.

CorpBench Score by model

Bars show score (0–100); markers show pass rate. Time is the average per scored run.

CorpBench Score, with pass-rate marker
  • 0.11¢ / task · 16s / task
  • 0.77¢ / task · 11s / task
  • 0.51¢ / task · 16s / task
  • 0.30¢ / task · 20s / task
0255075100CorpBench Score (0-100) and pass rate (%)

What this measures. CorpBench Work Score is a provisional composite combining average outcome score (60%), strict pass rate (25%), and efficiency against fixed, versioned baselines (15%). The marker shows pass rate across scored runs. Provider failures are excluded from scored runs and reported separately.

03Cost and speed

What each task costs in money and time

Model quality is only part of the deployment decision. This view compares outcome with the cost and time required to run each task.

Outcome against cost per task

Up is better work. Left is cheaper. The green corner highlights outcomes of 85+ at lower cost.

Outcome score (0-100), vertical axis

Smart and cheaper

Smart, but expensive

10092.58542.50
$0.0000$0.0015$0.0054$0.0152$0.0400

Cost per task (USD) · logarithmic scale

The vertical scale expands scores from 85–100 into the upper half of the chart.

Local runs appear at $0 in the cost view. This reflects API charges only; speed depends on the hardware used.

What this measures. Cost per task is total measured benchmark cost divided by scheduled runs. Time per task shows how long a task took, averaged across scored runs. Outcome score remains on the vertical axis in both views so cost and time are compared against the same result measure.

04Local models

How local models handle work

Compare workflow performance and speed on local hardware. No API or token charges.

Local model comparison

Published ranking, with the configuration behind each result.

  • Top local score

    Overall rank #8

    CorpBench Score

    84.0/ 100

    Pass rate
    77.0%
    Time / task
    19s
    Parameters
    8B
    Quantization
    Q4_K_M
    Run context
    32,768 tokens
    Runtime
    Ollama
    API cost / task
    FreeRuns on local hardware with no API or token charges. Hardware and electricity costs are not included.

  • Overall rank #16

    CorpBench Score

    62.4/ 100

    Pass rate
    38.0%
    Time / task
    45s
    Parameters
    9.7B
    Quantization
    Q4_K_M
    Run context
    32,768 tokens
    Runtime
    Ollama
    API cost / task
    FreeRuns on local hardware with no API or token charges. Hardware and electricity costs are not included.

What this measures. Scores and pass rates come from the same benchmark as the overall results. Speed depends on the machine and runtime used. Hardware and electricity costs are not included.

05Departments

Where the work gets harder

Overall scores can hide large differences between departments. A model that performs well across the benchmark may not be the best fit for every team.

Department outcome score

Average outcome score across scored runs in each department, on a 0 to 100 scale. Select a department to sort the models by its results.

Outcome score<6060+70+80+90+95+99+

Sorted by overall ranking

100
100
100
100
97
100
98
99
99
95
99
90
100
100
94
100
95
95
99
90

What this measures. Department outcome score is the average grader score across scored runs in that department. It includes partial credit, so it is not a pass rate or a recalculated CorpBench Score.

06Failure modes

How models fail at real work

A single score hides very different weaknesses. This view shows whether models stop early, skip required steps, give up, or get the handoff wrong.

Failure rate by behavior

Percentage of applicable scored runs exhibiting each failure. Lower is better. A run can show more than one, so the columns do not sum to 100%.

Failure rate0%≤1%≤3%≤5%≤10%≤20%>20%
Incomplete
workflow
Missing required
calls
Misordered required
calls
Abandoned
recoverable retry
Missing required
final state
Incorrect human
handoff
1%
1%
0%
0%
1%
0%
7.0%
6%
0%
4.5%
5%
0%
6.1%
6.1%
0%
0%
4.0%
1.0%
2.1%
2.1%
0%
0%
2.1%
0%

What this measures. Each value is the share of applicable scored runs in which the grader recorded that failure. Lower is better. A run can be recorded against more than one failure, so the columns overlap and do not sum to 100%. A blank cell means the failure was not scored for that model in this release.

07Reliability

How consistently each model succeeds

A strong average can hide inconsistent runs. Pass rate shows how often a model met every grading requirement.

Pass rate across scored runs

Share of scored runs that passed every grading requirement. Provider failures are excluded from scored runs and shown separately.

Pass rate
0%20%40%60%80%100%Pass rate (%)

What this measures. Pass rate is passed runs divided by scored runs. A run passes only when it meets every grading requirement. Provider failures after configured retries are excluded from scored runs and reported separately. Small score differences alone do not establish a meaningful capability difference.

08Methodology

How the results were produced

Each workflow runs in a fresh, deterministic workspace. Grading checks the final state, the required calls in order, and whether it completed or escalated correctly.

Workflows are observable work

Each workflow leaves evidence the grader can inspect: state changes, ordered tool calls, and a completion or escalation decision.

Grading is deterministic

The expected state and tool trace are fixed before the run. The same recorded evidence always receives the same grade.

Failures are separated

A provider failure after retries is excluded from scored runs and reported separately. Model errors, timeouts, and incorrect work remain in the scored results.

Repetitions define coverage

The release records workflow definitions and repetitions separately. A single run per workflow measures completion on that fixture; it does not establish consistency across variants.

About this release

Benchmark version
1.3
Workflows
100
Models
18
Published
Oct 18, 2026

Benchmark updates

Make better model choices

Get new benchmark results in your inbox as we test the latest models on real business workflows.

Unsubscribe anytime.

Build Capacity To Grow. Own Your Intelligence.

Bring the workflow slowing your sales, orders or delivery. We'll map the bottleneck and show you what the AI system could look like.

Book a working session