Are open-weight models fit for agentic workloads?

We compared widely used open-weight models to frontier closed ones. We find that open models trail on quality, fail unpredictably, and only tie on cost.

Pair the open-weight model Kimi K2.5 with ReAct, a simple tool-calling agent, and it handles 83% of telecom customer-service tasks. Pair the same Kimi with Claude Code, a more autonomous agent, and it handles zero. The model didn't change. The wrapper around it did.

This isn't a one-off. We tested DeepSeek V3.2 and Kimi K2.5, the two most widely deployed open-weight language models, against three closed-source models that span the frontier from two ends. Claude Opus 4.5 and Gemini 3 Pro set the quality bar. GPT-5.2 sets the cost bar. Together they answer two questions: how far is open-weight from the top, and is it really the cheap choice once you stop comparing against the premium frontier?

(Open-weight means the model's weights are published openly: anyone can download and run them. Closed-source models are only accessible through a vendor's API.) DeepSeek and Kimi see close to 10 million downloads per month on Hugging Face alone, and each downloaded copy can power many more local deployments. This isn't a comparison against fringe options; it's a comparison against what teams actually deploy. We ran every model with five agent designs across six benchmark categories that simulate realistic agent tasks.

How we tested

An agent is the software that wraps a language model with tools, rules, and a way to keep track of itself across many steps. The language model handles language. The agent handles everything that turns language into work.

We tested five different agent designs, spanning a spectrum:

We tested each model and agent combination across six of the most widely-used academic benchmarks for agent evaluation:

One hundred tasks per benchmark, scored by each benchmark's own evaluator. Full methodology is in the paper; all data and code are on the Open Agent Leaderboard.

On average, open-weight trails the frontier

Averaged across all five agent designs and all six categories of work, the ranking is clear:

Average success rate across all agent designs and tasks
Open-weight models trail the frontier closed-source models by 18 to 29 percentage points on average.
Closed-source
Open-weight
Each score is the average success rate across five agent designs and six benchmark suites (100 tasks each).

That's an 18- to 29-percentage-point gap between the open-weight models and the two frontier closed-source models. DeepSeek beats Kimi by 4 points, but neither comes near Gemini, let alone Opus. On raw average quality, the frontier still holds.

But the average is the wrong number to stop at.

Open-weight is more sensitive to agent design

The 18-to-29-point average doesn't reflect a uniformly worse model. It mixes a wider distribution: agent choice moves open-weight scores about twice as widely as it moves closed-source scores.

How much the score swings when you change the agent design
Best-vs-worst agent design for the same model, on the overall leaderboard score. Both open-weight models swing more than the closed-source frontier.
Closed-source
Open-weight
Each bar is the gap (in percentage points) between the best and worst agent design for the same model on the bench-weighted leaderboard score. Bigger means the model's score depends more on which wrapper you put around it.

For the closed-source models, the best and worst agent design land within 7 to 12 points. For the open-weight models, the same swing is 14 to 18 points.

If you're using Claude Opus or Gemini, picking any reasonable agent design keeps you within 7 to 12 points of the best score for that model. If you're using DeepSeek or Kimi, that range is 14 to 18 points, roughly twice as wide. Agent choice matters more for open-weight.

Closed-source models work with any reasonable agent design. Open-weight only works with the right one.

Some combinations are already competitive

That wider distribution has two ends. At the top end, specific (task × agent) combinations close most of the headline gap. Kimi with ReAct hits 83% on telecom technical support, within a point of Claude Opus's best on the same task. DeepSeek with ReAct reaches 82% on retail customer service.

In some combinations open-weight is already competitive. In others it's at zero. The average mixes both.

The natural follow-up: which combinations land where? You don't know in advance. Two specific failure shapes show up across the data.

Open-weight isn't worse everywhere. On the right combinations, it matches the frontier. But you can't tell which until you test.

Two kinds of performance sinks

The open-weight models show two distinct shapes of failure that the closed-source models don't. Both have the same practical effect: on a specific combination, the model essentially fails.

The first kind rules out a task category. Averaged across all five agent designs, both open-weight models land below 10% on personal-task automation. Every agent design we tried lands them at 16% or below.

A whole task category just doesn't work
Personal-task automation (AppWorld), averaged across all five agents.
score ≥ 20% sink (< 20%)
Outlined bars score below 20%.

The second kind shows up on a specific agent design. On telecom customer-service, the OpenAI agent SDK is the top-scoring wrapper for both closed-source models (84% with Opus, 89% with Gemini). It sinks both open-weight models to near zero on the same task (18% for DeepSeek, 0% for Kimi). The agent that lifts closed-source is the one that sinks open-weight on the same task. And nothing else in the data tells you where the next sink will land.

A specific agent design breaks open-weight
Telecom customer-service (TauBench-Telecom). The OpenAI agent SDK lifts closed-source, sinks open-weight.
score ≥ 20% sink (< 20%)
Outlined bars score below 20%.

The same agent that lifts closed-source to its best can break open-weight on the same task.

That is the deployment risk in concrete form. On the specific (task × agent) combinations you've measured, open-weight can be a strong choice. Move to one you haven't, and you have no basis to predict whether it will work. You find out at evaluation time.

On cost-effectiveness, open-weight is competitive, not dominant

The cost case for open-weight usually starts with DeepSeek versus Claude Opus. On the same autonomous agent, DeepSeek delivers about 62% of Opus's quality at roughly 2.7% of the cost: a 23× advantage on quality-per-dollar. That number is real but misleading. The real question isn't whether open-weight beats the most expensive frontier model. It's whether open-weight beats the cheapest closed-source option.

So we compared each model's best-quality configuration on cost and score, restricting closed-source to the cost-effective option:

Quality per dollar
Each model at its best-quality agent. Longer bars are better value.
Bar length is score divided by per-task cost.

At their best-quality operating points, the cost-effective closed-source model has higher quality and lower cost than either open-weight model. The "open-weight wins on cost" headline only survives when you compare to Opus, where the gap on quality-per-dollar is 15× or more. Against Gemini the gap is smaller (a few-fold, depending on configuration); against the cheap end of closed-source, it closes entirely.

DeepSeek retains one specific niche: at absolute minimum cost with no quality floor, it drops to $0.09/task with the OpenAI agent SDK, about half the cheapest closed-source configuration. But quality there is 32%, which doesn't beat the cheap closed-source option on quality-per-dollar either. You're just running a worse model for fewer cents.

Even on price, open-weight isn't better than cost-effective closed-source models. But it's competitive.

What this means for deployment

For a (task × agent) combination you've measured directly, open-weight can be a strong choice. DeepSeek already reaches 80%+ on its best combinations, at a fraction of premium-frontier prices. If you control the workload and can run your own evaluation on the exact combination you plan to ship, you can lock in the cost gap: meaningful against Opus and Gemini, modest against GPT-5.2.

For tasks or agent designs you haven't measured, open-weight is not safe to deploy. The 18-to-29-point gap to the performance frontier matters, but the bigger problem is unpredictability. Open-weight scores swing twice as widely as closed-source scores under agent choice, and the combinations that collapse only reveal themselves at evaluation time.

For general-purpose agents meant to handle whatever comes their way, open-weight is not ready. A general-purpose system meets combinations it hasn't been measured on and runs inside diverse agent designs. Both are the conditions where open-weight is hardest to predict.

What's next

This is a snapshot. New open-weight models ship constantly, and the picture will shift with each release cycle. The structure has held across every open-weight model we've tested so far: competitive on the tasks and agents they appear tuned for, unpredictable elsewhere. The coming releases will tell us how durable that pattern is.

We'll keep the leaderboard current as new models arrive. If you've run your own evaluations on different models, agents, or benchmarks, open a PR on GitHub and we'll add them.