AsyncTool
Agents coordinate several tool-use tasks while tool responses arrive after simulated delays. The evaluator checks function calls, parameters, task completion, and the final environment state. Scheduling metrics measure whether agents use waiting time effectively.
The benchmark reports several correctness and efficiency measures rather than one interchangeable score.
- Released
- May 2026
- Kind
- Function calling
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- APIs and tools
- Items from
- Generated
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Agents
- Strongest model tested
- GPT-5 #71 on the capability index
- Last model added
- September 2025
- Scores refreshed
- November 20th, 2026
In the authors’ words
Large language model (LLM)-based agents have shown strong capabilities in using external tools to solve complex tasks. However, existing evaluations often overlook the temporal dimension of tool use, especially the impact of tool response latency, and are usually limited to single-task settings. In real-world applications, multiple tasks often need to be executed concurrently, and overall efficiency depends on whether an agent can use idle time while waiting for tool responses. We refer to this capability as asynchronous tool calling. To evaluate it, we propose AsyncTool, a benchmark for assessing LLM-based agents in interactive multi-task tool-use environments with delayed tool feedback. AsyncTool presents multiple heterogeneous tasks simultaneously and simulates realistic tool response latency during execution. Using a hybrid data evolution strategy, we construct a diverse asynchronous multitasking dataset that covers multiple scenarios and tool-use patterns. We evaluate models at the step, sub-task, and task levels, and introduce efficiency-oriented metrics to measure task coordination and completion efficiency. Extensive experiments show that delayed tool feedback poses substantial challenges to current agents and leads to clear performance degradation. Models that better coordinate task switching, dependency tracking, and state maintenance achieve stronger performance on AsyncTool. Our analysis identifies key failure modes of current tool-using agents and provides practical insights for designing future systems with stronger temporal reasoning and coordination capabilities.
GPT-4.1 leads with 38.1% among models released by September 2025 GPT-4.1 leads with 38.1% among models released by September 2025
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in September 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 38.1 GPT-4.1
- 35.5 GPT-4o (snapshot unspecified)
- 35.3 Gemini 2.5 Pro (snapshot unspecified)
- 34.5 Qwen2.5-72B-Ins.
- 34.4 Qwen-Max (snapshot unspecified)
- 31.3 GPT-5
- 28.9
DeepSeek-V3.1-Terminus
- 26.3
Qwen3-30B-A3B-Instruct
- 24.9 Qwen2.5-32B-Ins.
- 24.4 Kimi K2
- 21.9 Qwen2.5-14B-Ins.
- 20.1
Qwen3-32B
- 18.8
Qwen3-14B
- 15.2 GLM-4-32B
- 14.0
Qwen3-8B
- 8.43 Qwen2.5-7B-Ins.
- 6.88 GLM-4-9B-Chat
- 6.74 LLaMA-3.1-8B-Ins.
- 5.34 LLaMA-3.3-70B-Ins.
- 3.37 LLaMA-3.1-70B-Ins.
- 1 GPT-4.1 38.1%
- 2 GPT-4o (snapshot unspecified) Random zero-to-one-turn delay 35.5%
- 3 Gemini 2.5 Pro (snapshot unspecified) Random zero-to-one-turn delay 35.3%
- 4 Qwen2.5-72B-Ins. Few-shot 34.5%
- 5 Qwen-Max (snapshot unspecified) Random zero-to-one-turn delay 34.4%
- 6 GPT-5 31.3%
- 7
DeepSeek-V3.1-Terminus 28.9%
- 8
Qwen3-30B-A3B-Instruct Jul 2025, Random zero-to-one-turn delay 26.3%
- 9 Qwen2.5-32B-Ins. 24.9%
- 10 Kimi K2 Sep 2025 24.4%
- 11 Qwen2.5-14B-Ins. Few-shot 21.9%
- 12
Qwen3-32B Random zero-to-one-turn delay 20.1%
- 13
Qwen3-14B 18.8%
- 14 GLM-4-32B 15.2%
- 15
Qwen3-8B Random zero-to-one-turn delay 14%
- 16 Qwen2.5-7B-Ins. Random zero-to-one-turn delay 8.43%
- 17 GLM-4-9B-Chat 6.88%
- 18 LLaMA-3.1-8B-Ins. Few-shot 6.74%
- 19 LLaMA-3.3-70B-Ins. 5.34%
- 20 LLaMA-3.1-70B-Ins. Random zero-to-one-turn delay 3.37%
Source: The AsyncTool paper
- 1 GPT-4.1 38.1%
- 2 GPT-5 31.3%
- 3
DeepSeek-V3.1-Terminus 28.9%
- 4
Qwen3-30B-A3B-Instruct Jul 2025, Random zero-to-one-turn delay 26.3%
- 5 Kimi K2 Sep 2025 24.4%
- 6
Qwen3-32B Random zero-to-one-turn delay 20.1%
- 7
Qwen3-14B 18.8%
- 8
Qwen3-8B Random zero-to-one-turn delay 14%
- 9 GLM-4-9B-Chat 6.88%
Source: The AsyncTool paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-4.1 | 38.1% | Apr 2025 | Official leaderboard | |
| 2 | GPT-4o (snapshot unspecified) closed weights | 35.5% | Random zero-to-one-turn delay | Official leaderboard | |
| 3 | Gemini 2.5 Pro (snapshot unspecified) closed weights | 35.3% | Random zero-to-one-turn delay | Official leaderboard | |
| 4 | Qwen2.5-72B-Ins. open weights | 34.5% | Few-shot | Official leaderboard | |
| 5 | Qwen-Max (snapshot unspecified) closed weights | 34.4% | Random zero-to-one-turn delay | Official leaderboard | |
| 6 | GPT-5 | 31.3% | Aug 2025 | Official leaderboard | |
| 7 | | 28.9% | Sep 2025 | Official leaderboard | |
| 8 | | 26.3% | Random zero-to-one-turn delay | Jul 2025 | Official leaderboard |
| 9 | Qwen2.5-32B-Ins. open weights | 24.9% | Official leaderboard | ||
| 10 | Kimi K2 Sep 2025 | 24.4% | Sep 2025 | Official leaderboard | |
| 11 | Qwen2.5-14B-Ins. open weights | 21.9% | Few-shot | Official leaderboard | |
| 12 | | 20.1% | Random zero-to-one-turn delay | Apr 2025 | Official leaderboard |
| 13 | | 18.8% | Apr 2025 | Official leaderboard | |
| 14 | GLM-4-32B open weights | 15.2% | Official leaderboard | ||
| 15 | | 14% | Random zero-to-one-turn delay | Apr 2025 | Official leaderboard |
| 16 | Qwen2.5-7B-Ins. open weights | 8.43% | Random zero-to-one-turn delay | Official leaderboard | |
| 17 | GLM-4-9B-Chat | 6.88% | Jun 2024 | Official leaderboard | |
| 18 | LLaMA-3.1-8B-Ins. open weights | 6.74% | Few-shot | Official leaderboard | |
| 19 | LLaMA-3.3-70B-Ins. open weights | 5.34% | Official leaderboard | ||
| 20 | LLaMA-3.1-70B-Ins. open weights | 3.37% | Random zero-to-one-turn delay | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-4.1 | 38.1% | Apr 2025 | Official leaderboard | |
| 2 | GPT-5 | 31.3% | Aug 2025 | Official leaderboard | |
| 3 | | 28.9% | Sep 2025 | Official leaderboard | |
| 4 | | 26.3% | Random zero-to-one-turn delay | Jul 2025 | Official leaderboard |
| 5 | Kimi K2 Sep 2025 | 24.4% | Sep 2025 | Official leaderboard | |
| 6 | | 20.1% | Random zero-to-one-turn delay | Apr 2025 | Official leaderboard |
| 7 | | 18.8% | Apr 2025 | Official leaderboard | |
| 8 | | 14% | Random zero-to-one-turn delay | Apr 2025 | Official leaderboard |
| 9 | GLM-4-9B-Chat | 6.88% | Jun 2024 | Official leaderboard |
Rankings by part
We read 4 parts from the benchmark’s paper, each ranking models on one part of the test, such as SIMILAR#2, SIMILAR#3 and CROSS#2. GPT-4.1 leads 3 of them, and GPT-5 leads the other one. Arrows show how far a model moved from its overall rank among the models we track.
SIMILAR#2
9 models we track have a score here.
- 1 GPT-4.1 50%
- 2 GPT-5 41.7%
- 3 Kimi K2 Sep 2025 ↑2 33.3%
- 4
DeepSeek-V3.1-Terminus ↓1 32.5%
- 5
Qwen3-32B ↑1 31.7%
- 6
Qwen3-30B-A3B-Instruct Jul 2025 ↓2 30%
- 7
Qwen3-14B 27.5%
- 8
Qwen3-8B 25.8%
- 9 GLM-4-9B-Chat 12.5%
SIMILAR#3
9 models we track have a score here.
- 1 GPT-4.1 36.7%
- 2
DeepSeek-V3.1-Terminus ↑1 24.6%
- 3 GPT-5 ↓1 21.7%
- 4 Kimi K2 Sep 2025 ↑1 20.8%
- 5
Qwen3-30B-A3B-Instruct Jul 2025 ↓1 15.4%
- 6
Qwen3-14B ↑1 14.6%
- 7
Qwen3-32B ↓1 13.3%
- 8 GLM-4-9B-Chat ↑1 7.08%
- 9
Qwen3-8B ↓1 5.42%
CROSS#2
9 models we track have a score here.
- 1 GPT-4.1 47.7%
- 2 GPT-5 43.9%
- 3
DeepSeek-V3.1-Terminus 35.6%
- 4
Qwen3-30B-A3B-Instruct Jul 2025 32.6%
- 5 Kimi K2 Sep 2025 30.3%
- 6
Qwen3-14B ↑1 28%
- 7
Qwen3-32B ↓1 26.5%
- 8
Qwen3-8B 15.9%
- 9 GLM-4-9B-Chat 8.33%
CROSS#3
9 models we track have a score here.
- 1 GPT-5 ↑1 28.6%
- 2
DeepSeek-V3.1-Terminus ↑1 27.7%
- 3 GPT-4.1 ↓2 27.3%
- 4 Kimi K2 Sep 2025 ↑1 20%
- 5
Qwen3-30B-A3B-Instruct Jul 2025 ↓1 16.8%
- 6
Qwen3-32B 14.1%
- 7
Qwen3-14B 13.2%
- 8
Qwen3-8B 5.00%
- 9 GLM-4-9B-Chat 2.73%
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: The AsyncTool paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-4.1 | 38.1% | ||||
| GPT-4.1 | Random zero-to-one-turn delay | 34.5% | |||
| GPT-5 | 31.3% | ||||
| GPT-4.1 | Random one-to-two-turn delay | 29.5% | |||
| 28.9% | |||||
| GPT-4.1 | Two-turn delay | 26.4% | |||
| Random zero-to-one-turn delay | 26.3% | ||||
| Random zero-to-one-turn delay | 26.3% | ||||
| Kimi K2 Sep 2025 | 24.4% | ||||
| 21.5% | |||||
| Random one-to-two-turn delay | 20.9% | ||||
| Two-turn delay | 20.6% | ||||
| Random zero-to-one-turn delay | 20.1% | ||||
| 19.1% | |||||
| 18.8% | |||||
| Random one-to-two-turn delay | 16.6% | ||||
| Two-turn delay | 14.6% | ||||
| Random zero-to-one-turn delay | 14% | ||||
| Random one-to-two-turn delay | 11.2% | ||||
| Few-shot | 11.2% | ||||
| Two-turn delay | 11% | ||||
| 10.7% | |||||
| Random one-to-two-turn delay | 7.16% | ||||
| GLM-4-9B-Chat | 6.88% | ||||
| Two-turn delay | 6.60% |
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: The AsyncTool paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-4.1 | 38.1% | ||||
| GPT-4.1 | Random zero-to-one-turn delay | 34.5% | |||
| GPT-5 | 31.3% | ||||
| GPT-4.1 | Random one-to-two-turn delay | 29.5% | |||
| 28.9% | |||||
| GPT-4.1 | Two-turn delay | 26.4% | |||
| Random zero-to-one-turn delay | 26.3% | ||||
| Random zero-to-one-turn delay | 26.3% | ||||
| Kimi K2 Sep 2025 | 24.4% | ||||
| 21.5% | |||||
| Random one-to-two-turn delay | 20.9% | ||||
| Two-turn delay | 20.6% | ||||
| Random zero-to-one-turn delay | 20.1% | ||||
| 19.1% | |||||
| 18.8% | |||||
| Random one-to-two-turn delay | 16.6% | ||||
| Two-turn delay | 14.6% | ||||
| Random zero-to-one-turn delay | 14% | ||||
| Random one-to-two-turn delay | 11.2% | ||||
| Few-shot | 11.2% | ||||
| Two-turn delay | 11% | ||||
| 10.7% | |||||
| Random one-to-two-turn delay | 7.16% | ||||
| GLM-4-9B-Chat | 6.88% | ||||
| Two-turn delay | 6.60% |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The AsyncTool paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official AsyncTool leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Tool use
- AutomationBench In use Agent completes end-to-end business workflows inside a simulated company exposing dozens of REST APIs. Graded on resulting state.
Gemini 4 Neon 56.3%
- MCP Atlas In use Multi-step workflows over real MCP servers. Scores tool-use competency on realistic tasks.
Gemini 4 Argon 90.7%
- BFCL V4 In use Models call functions and tools, including multi-turn agentic tasks with web search and memory. Checked against expected calls and outcomes.
Qwen3.7-Max 75%
- τ³-bench Banking In use Agents act as banking customer-service reps over many turns with a simulated user, following policy. Pass^1 on the final database state.
Qwen 3.8 Max 55.1%
- Toolathlon-Verified In use 108 long-horizon tasks across 32 apps through about 600 MCP tools. Execution-based checkers inspect final environment state. GLM-5.3-Flash 78.4%
- DeepSearchQA In use Questions that need multi-step search and synthesis. Final answers graded against references. Kimi K3 95%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 14 months ago, in September 2025, for DeepSeek-V3.1-Terminus. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (20 entries).
- Pass Models scored 9 models we track, plus 11 other entries on its leaderboard.
- Warning Newest score The newest came 14 months ago, in September 2025, for DeepSeek-V3.1-Terminus.
- Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5 (#71).
- Pass Headroom The best is 38.1%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 25 rows for models we track and 11 other entries at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.