AIM Enterprise: Agentic Enterprise Benchmark
Enterprises use LLMs everyday for their regular tasks. To find most cost efficient LLMs, we designed AIM Enterprise, an agentic enterprise benchmark, where we used 69 enterprise tasks in different categories.
Benchmark results
Each model delivered one file per task. Scores are relative: the judges rank every answer against the other answers to the same task, so the average score is 50 by design. A 55 means a model’s answers ranked above most others, not that 55% of its work was correct.
Two models finished clearly ahead. Claude Opus 5 scored 66.4 and GPT 5.6 Sol 62.4, more than 7 points above the third-placed setup. Their lead does not depend on which tasks we picked: we rescored the leaderboard 4,000 times on random re-selections of the 69 tasks, and those two held first and second place every time.
The next nine models scored between 51.4 and 55.3. Sixty-nine tasks cannot separate scores that close, so their order changes depending on which tasks are counted. Kimi K3 is the exception in that group, finishing above Claude Sonnet 5 in 94% of the rescorings.
Opus 5 also won the hardest tasks. We took the 17 tasks with the lowest average score across all 16 setups, fixed by that rule before anyone’s standing on them was calculated, and Opus 5 beat all 15 other setups on that subset.
Opus 5 has no weak tasks. Its best score, 80.6, is the highest anyone posted, and its worst, 52.9, still beats an average answer. It scored above 60 on 62 of the 69 tasks and finished first on 43 of them. GPT 5.6 Sol won 13 tasks and no other model won more than four.
The two scoring models disagreed about which of the leaders should come first. Each put its own company’s model on top, so the published order combines both rankings.
Cost and score
Cost covers the 11 setups billed per run. Five setups, including both leaders, ran on subscriptions that record no per-run charge.
The best-scoring priced setup is also the cheapest. GPT 5.6 Luna scored 55.3 at $0.032 per task, just under DeepSeek V4 Flash at $0.033. Claude Sonnet 5 scored 53.2 at $1.46, which is 46 times the cost and 2.1 points lower.
Paying more buys little here. Across the 11 priced setups the correlation between cost per task and score is 0.48. Kimi K3 at $0.92 and Qwen 3.8 Max at $0.74 scored in the same range as Luna at $0.03.
A fast run is a warning sign, but a slow one proves nothing. Across all 16 setups the correlation between time per task and score is 0.55, and the four fastest were four of the five lowest scorers. Inkling Small finished a task in 49 seconds and came second from bottom. Among the nine models bunched in the middle, which took between 127 and 569 seconds, that correlation drops to −0.06.
Costs are what the harness billed at its bundled price list, not a quote. Provider list prices are in our LLM pricing comparison. Failed attempts count in a model’s total but not in its per-task figure, and the two differ most for Grok 4.5: $26.14 spent against $22.84 of delivering runs.
Judge disagreement
Two models scored every file and they often disagreed. On about a third of the individual scores the two landed more than a quarter of the scale apart, and nobody has reviewed those rows.
They agreed on the ends and not the middle. Both put Opus 5 and GPT 5.6 Sol above everyone else, but 11 of the 16 setups land in a different position depending on which scoring model you follow.
Every task ran once and was scored once. A model that produced no file was run again, but no file was ever scored twice, so no result came from picking the better of two attempts.
A seventieth task, an inquiry-handling playbook, is left out of every figure here. Its input file was missing at run time and only one setup ever produced a file for it.
AIM Marketing benchmark
The same method runs on marketing work: finding offerings a competitor publishes and AIMultiple does not, building a best-fit account list, and producing a personalized sales deck. Each task scores 0 to 100 and the overall score is the mean of the three. A website reputation audit runs alongside them and is reported on its own axes, because it has no fixed maximum score.
Full results: the agentic marketing benchmark.
AIM IT benchmark
Twelve models were told to invent a benchmark, build it, and run four models through it, twice each. None of the 24 attempts passed every criterion, and six of the rubric’s checks were passed by none of them. Claude Opus 5 led on text-to-SQL at 78.2 and Kimi K3 on tool calling at 74.3.
Full results: can LLMs design a benchmark.
AIM VC benchmark
Thirteen models were asked to name a company’s clients with dated evidence, across three target companies. Claude Opus 5 scored 89.1 of 100 and Claude Fable 5 85.0, and those two were the only ones that stayed above 76 on all three targets. Most of the others collapsed on the target whose clients are named in podcast sponsor reads rather than indexed pages.
Full results: the client-list benchmark.
Add as preferred sourceMethodology
The 69 tasks were written by AIMultiple’s founder against real company decisions and mapped to APQC Process Classification Framework ids.1
They cover strategy (16 tasks), marketing (15), HR (7), sales (6), operations (5), IT and finance (4 each), and seven smaller areas. Each task names its deliverable: a results.csv with a fixed column list, typically 10 rows and 8 columns, with an explicit rule for every integer field.
How the models ran
Thirteen models ran in 16 setups, a setup being one model under one agent program. Eleven ran on opencode 1.15.13 through OpenRouter. The rest ran on the agent programs their own vendors ship, Claude Code and Codex, billed against subscriptions.
Every setup got the same frozen prompt, live web access through a scraping API and a two-hour limit. Prompts were never tuned per model, and the scoring rubrics never reached the machine that ran the tasks.
That produced 1,104 files, one per model per task. Six opencode setups missed 25 cells on the first pass because the agent’s file search walked paths its own sandbox then refused to open, which stalled the run. Rerunning those 25 in an isolated directory recovered all of them, so delivery is reported as both a first-pass and a final rate. Counting every restart, 83 of the 759 cells with usage records took more than one try.
How the checks and judges work
Deterministic checks run first and no judge sees a file that fails them: column set and row count, RFC 4180 parsing,2
integer formatting, no empty cells, no duplicate rows, and sort order where the task requires one.
A file that fails scores zero rather than being dropped, because a file nobody can parse is a result. DeepSeek V4 Flash failed on 6 of its 69 files, Inkling Small on 2, MiniMax M3 and GPT 5.6 Terra on one each. Removing every task where any setup failed leaves the two leaders and the overall shape intact.
The 1,094 files that passed went to two judges: GPT 5.6 Sol through the Codex CLI and Claude Opus 5 through the Claude Code CLI, both at high reasoning effort with live web access.
Each column of the deliverable goes to its own subagent, which sees that column’s answers from every model and nothing else, under anonymous row ids shuffled separately for each judge. The judge ranks those answers against each other and cannot call any two equal. The two rankings are then combined by adding up each answer’s position under each judge. That produced 90,530 scores.
Anonymization is not cosmetic. Measured with model names visible, Sol ranked GPT answers 8.4 percentiles above where Opus ranked them, and Opus ranked Anthropic answers 4.9 percentiles above where Sol ranked them. It does not remove the effect: the runs behind this leaderboard were anonymized and each judge still put its own company’s setup first.
How scoring works
A model’s score is the sum of its column averages, rescaled so the highest possible total is 100. That ceiling depends on how many models passed the checks, which is why these scores compare inside this benchmark and nowhere else.
To test how much the ranking depends on the task selection, we rescored the leaderboard 4,000 times, each time on a random re-selection of the 69 tasks. That measures task sensitivity only. It does not measure how much a rerun of the same task would move a score, because no task was scored twice.
The hardest quarter is the 17 tasks with the lowest average score across all 16 setups. That rule was fixed before any model’s standing on those tasks was calculated.
Twelve of the 13 models here also run in the other benchmarks above, so their ordering carries across the series. The numbers do not, because each benchmark sets its own scale.
This benchmark cannot be trimmed the way the others can. Every score is a position relative to the models it ran against, so dropping one setup would rescore all the rest instead of removing a bar. Qwen 3.8 Max runs this benchmark only and stays in the chart for that reason.
FAQs
Not on this evidence. The scale is relative, so even the top score of 66.4 says only that a model’s answers ranked above the others, and the rows where the two scoring models substantially disagree are unreviewed. Vendors building this kind of agent are listed in our enterprise AI companies breakdown, and AIMultiple automates processes like these.
Because nine of them are genuinely close, and 69 tasks cannot separate differences under about 1.5 points. The benchmark tells strong models from weak ones, not one mid-table model from the next.
Not for output quality, in the three cases we could test. Three models each ran under two agent programs and no pair differed by more than 0.83 points. The difference showed up in operations instead: only the opencode setups lost cells to a sandbox conflict, and only the subscription-billed programs left no usage record.
Cite this research
Pick the format that matches where you're publishing. Pasting the link version into your CMS preserves the backlink.
@misc{kalelioglu2026,
author = {Kalelioğlu, Berk},
title = {{AIM Enterprise: Agentic Enterprise Benchmark}},
year = {2026},
month = aug,
howpublished = {\url{https://aimultiple.com/agentic-enterprise}},
note = {AIMultiple. Retrieved August 14, 2026}
}
Be the first to comment
Your email address will not be published. All fields are required. Comments are left in their original language.