Agent Evaluation · Economic Grounding

Faster. Cheaper.
Not good enough yet.

First surfaced in Tandemly Briefing — 2026-07-30.

Baidu researchers built a benchmark where every office task carries an economic price tag: human labor time and cost. Every model they tested completed tasks faster and cheaper than a human worker. None produced output at human deliverable quality. The gap between "cheaper than a human" and "replaces a human" is real, measurable, and still wide.

Core concept
Economic grounding: each benchmark task carries a human labor time estimate and a cost proxy, so agent output quality can be priced against the human alternative rather than ranked only against other models.
scroll to explore

Pass/fail misses
the point.

Standard agent benchmarks ask whether a task was completed. That question made sense when agents were unreliable enough that completion itself was the goal. Agents are now reliable enough that "did it finish" is no longer the useful question.

Most agent benchmarks were designed when the dominant failure mode was an agent that could not complete a task at all. A model that gets stuck in a loop, generates nonsense, or simply gives up does not complete the task, and task completion rate captures that failure cleanly.

But frontier models now complete a large fraction of well-defined tasks on the first try. The leaderboards reflect this: completion rates are high, scores converge, and the gaps between models on standard benchmarks are often smaller than the gap between any model and the upper bound of what the benchmark rewards. The metric is still moving, but it is moving more slowly, and the improvements it measures are increasingly incremental.

The more pressing question for anyone deploying agents on professional work is not whether the task was completed, but whether the completed output is actually usable. A research summary that needs two hours of heavy editing before it can be sent to a client is not an equivalent substitute for one written by a skilled analyst. A formatted document with structural errors is not the same as a clean deliverable. Completing the task and producing a usable deliverable are not the same thing, and the difference between them has an economic cost that standard benchmarks do not capture.

The question this paper asks

What happens when you replace "did the agent complete the task" with "is the agent's output worth what it would cost a human to produce the same quality"? Does the economic case for agent deployment hold up under that framing?

Office tasks, priced
against human labor.

OmegaUse-OfficeVal is a benchmark of long-horizon office-suite tasks, each annotated with a human labor time estimate and a cost proxy. The benchmark enables two metrics standard evaluations omit: quality relative to human output, and quality per dollar of agent cost.

The tasks are the kind of work that shows up in professional office environments: documents that require multiple synthesis steps, formatted outputs with structured requirements, tasks where the quality of the final product is visible and assessable. "Long-horizon" means each task requires multiple intermediate decisions before the output is complete, not a single-step answer.

Each task carries two economic tags. The first is a human labor time estimate: how long a skilled worker would need to produce an acceptable deliverable for that task, based on professional time standards. The second is a derived cost figure drawn from labor rates for that kind of work. Together, they give a concrete human-cost baseline for each task in the benchmark.

With that baseline in hand, the benchmark computes two things for any agent run. Quality is measured by a rubric applied to the output deliverable, not just to task completion. Cost is measured as total agent spend, API calls included. Quality-per-dollar for the agent is then compared against the quality-per-dollar from a human worker on the same task, using the economic tag as the human data point.

Standard benchmark
Did the agent complete the task? One score per task, measured against a correct answer or a completion threshold. Agents that finish faster and cheaper look strictly better, regardless of output quality.
OmegaUse-OfficeVal
Is the output worth what it cost? Each task is tagged with human labor time and cost. Agent quality is scored on the deliverable. Quality-per-dollar compares agent and human on the same denominator.
Why the denominator matters

Most cost-aware agent research measures how much the agent spends (API tokens, latency) relative to task success rate. OmegaUse-OfficeVal measures something different: how much value the agent produces relative to the human labor it would replace. The two questions are related but not the same. An agent that costs $0.08 per task where a human costs $40 per task looks extremely cheap. If the agent's output quality requires $30 in human correction, the agent costs $38 per usable deliverable, not $0.08.

Cheaper. Faster.
Still not there.

Every model tested completed tasks faster than a human worker and at lower direct agent cost. None produced human-quality deliverables on the long-horizon tasks. The economic case for full automation did not close.

All models
100%
beat humans on speed
All models
100%
beat humans on direct cost
Deliverable quality
0
models reached human parity
Finding 1: The speed-and-cost advantage is real but incomplete

Models complete the tasks. They do so faster than a human worker and at lower direct API cost. On those two dimensions, the benchmark results look like a success story for agent deployment.

The incomplete part is that "completing the task" and "producing a deliverable that does not require substantive revision" are not the same thing for long-horizon professional work. A task that is technically complete but requires an hour of human editing to become usable is not a time-and-cost win. The benchmark makes this distinction explicit by scoring the output quality separately from task completion.

Finding 2: No model closed the quality gap

None of the models tested produced outputs that matched human deliverable quality on the long-horizon office tasks. The gap was not marginal. Outputs required substantive review before they were ready to use as professional products.

This does not mean agents are bad at office work in general. The specific failure mode here is long-horizon tasks where quality is assessed on the deliverable structure and professional form, not just on factual content. Shorter, more structured tasks may tell a different story. But for the class of work where professional output quality is the standard, the economic case for full replacement, rather than assistance, does not yet hold.

Finding 3: The honest accounting inverts the headline

The common framing of agent adoption as "faster and cheaper than hiring" relies on an implicit assumption: that the agent's output is usable as delivered. OmegaUse-OfficeVal shows that assumption fails for long-horizon office tasks at current capability levels. When the cost of human review and correction is added to the agent's direct cost, the economic advantage narrows or disappears, depending on the task and the required quality level.

The benchmark also surfaces a useful diagnostic: the tasks where agents score highest on speed and direct cost are not necessarily the tasks where they produce the strongest deliverables. The correlation between "fast and cheap" and "high quality" is not assumed in the benchmark design, and the results suggest it should not be assumed in deployment decisions either.

Scope and limitations

The benchmark reflects Baidu's definition of office-suite tasks and their labor cost estimates. Different organizations, task definitions, and quality rubrics would produce different numbers. The human labor time estimates are derived from professional time standards, not measured from a controlled comparison group. The results describe what happened at one point in model capability development and will change as models improve. The core contribution is the methodology: attaching economic tags to benchmark tasks so the comparison is possible, not a fixed claim about which models close the gap.

What this means
for how you evaluate.

The benchmark design is the contribution, not just the result. The question OmegaUse-OfficeVal asks, "is the agent worth what it would cost a human to do the same quality work," is a question any team can apply to their own evaluation setup with modest additional effort.

1
For teams evaluating agents for professional task automation
Add quality-per-dollar to your evaluation metrics alongside task completion rate. The two metrics can diverge significantly. An agent that completes 90% of tasks at one-tenth the direct cost may still require 40% of a human's effort in review and correction, which changes the economics materially. Track review time as a cost, not just a process step.
2
For benchmark designers
The gap that OmegaUse-OfficeVal surfaces, between task completion and deliverable quality, is structural, not incidental. Any benchmark evaluating agents on professional work should include a quality rubric applied to the output, not just a completion signal. Attaching labor-time estimates to tasks turns quality-per-dollar into a calculable metric that anchors agent performance to something outside the benchmark itself.
3
For decision-makers comparing vendors or models
Leaderboard rankings compare agents against each other. The question for a deployment decision is different: compared to the human labor the agent would replace, does it produce equivalent value at lower total cost once review is included? That comparison requires a human baseline with a realistic cost figure. An agent that ranks first on an all-agent leaderboard may still not clear the human-replacement bar on your specific task class.
4
For teams already deploying agents in office workflows
This research is an invitation to audit what your team actually does with agent output. If reviewers are routinely making substantive edits before outputs go out, the agent is providing a first draft, not a deliverable. That is a legitimate and often valuable use case, but it should be measured honestly. Track the delta between what the agent produces and what actually gets sent, and price it.

Where to go
from here.

The benchmark methodology is the most portable takeaway from this paper. Several of the steps below can be applied to an existing evaluation setup without waiting for a new benchmark release.

1
Read the paper
Zhou, Zhao et al. (2026). OmegaUse-OfficeVal: Benchmarking LLM Agents on Long-Horizon Office-Suite Tasks with Economic Grounding. Baidu. arXiv:2607.27155.
2
Add labor-time estimates to your existing eval tasks
Take five tasks from your current agent evaluation set and estimate, honestly, how long a skilled professional would need to produce equivalent-quality output. Write the numbers down. That estimate is your human cost baseline and it turns your pass/fail rate into a quality-per-dollar comparison.
3
Track review time as a cost line item
For a sample of agent outputs that go through human review before use, time the review step. Even rough estimates, averaged over ten to twenty tasks, reveal whether the agent is producing deliverables or drafts. The difference determines which class of economic argument applies to the deployment.
4
Report quality-per-dollar in any comparative evaluation
When comparing agent options internally, compute total cost per usable output (direct API cost plus estimated review cost) and divide by your quality score. Present this as a primary metric alongside accuracy and speed. It is the one metric that captures both sides of the economic equation at once.
5
Pair with the Token Reduction cost study for a fuller picture
OmegaUse-OfficeVal measures value relative to human labor. The companion research on token reduction and API billing measures agent cost more precisely, showing how direct API spend can diverge from intuitive estimates. Together, they frame the economics from both ends: what the agent costs to run, and whether what it produces is worth that cost.