Last verified
Agents and computer useVendor-reportedPrices verified

AutomationBench leaderboard

As of November 8, 2026, MiniMax M3.1 Flash (MiniMax) leads AutomationBench at 62.8%, ahead of DeepSeek V4.1 Flash at 60.5% and MiMo-V2.6-Pro at 53.1%. 23 systems ranked. Every score here is published by the model's own lab; each row links to the page it comes from.

See the ranking →All benchmarks →Model prices →
Top score
62.8%
MiniMax M3.1 Flash
Systems ranked
23
across 3 settings
Independent rows
0/26
all vendor-reported
Cheapest near top
$0.0842/1M
MiniMax M3.1 Flash · within 5 pts
With API prices
23/23
rates from our price index
Newest score
Oct 19
2026 · newest result

What is AutomationBench? AutomationBench tests whether an AI agent can carry out a business workflow across several apps through their APIs, finding the endpoints itself and following company rules. Made by Zapier, April 2026.

§ 01 / LEADERBOARD

Top models on AutomationBench, November 2026

20 entries · Main set (as reported)
# Model Score $ / task Price / 1M Context Source
1 MiniMax M3.1 Flash MiniMax
62.8%
v1.0.6, 600 public tasks 		— $0.15 / $0.6 1M Vendor-reported MiniMax
2 DeepSeek V4.1 Flash DeepSeek · listed as “DeepSeek V4.1 Flash (1017)”
60.5%
v1.0.6 		— $0.3 / $1.2 1M Vendor-reported DeepSeek
3 MiMo-V2.6-Pro Xiaomi MiMo
53.1%
v1.0.6 		— $0.435 / $0.87 1M Vendor-reported Xiaomi
4 MiMo-V2.6-Flash Xiaomi MiMo
52.3%
v1.0.6 		— $0.14 / $0.28 1M Vendor-reported Xiaomi
5 Gemini 4 Argon Google DeepMind
51.3%
as reported 		— $2 / $10 1M Vendor-reported Google DeepMind
6 GLM-5.3 Zhipu (Z.ai / GLM)
48.2%
v1.0.6 		— $1.4 / $4.4 1M Vendor-reported Z.ai (Zhipu)
7 DeepSeek V4 Pro DeepSeek
43.2%
pass@1 		— $1.32 / $3.96 1M Vendor-reported DeepSeek
8 GPT-6 Astra OpenAI
41.4%
max effort, v1.0.6 		$1.73 $10 / $50 1.05M Vendor-reported OpenAI
9 Claude Opus 5.5 Anthropic
40%
run by Zapier, no fallback models 		— $4 / $20 1M Vendor-reported Anthropic
10 DeepSeek V4 Flash Deprecated DeepSeek
37.7%
pass@1 		— $0.44 / $1.32 1M Vendor-reported DeepSeek
11 GPT-6.1 Sol OpenAI
36.1%
max effort, v1.0.6 		$0.30 $2 / $10 1.05M Vendor-reported OpenAI
12 GPT-6 Sol OpenAI
33.2%
xhigh effort, v1.0.6 		$0.27 $2 / $10 1.05M Vendor-reported OpenAI
13 Claude Fable 5.1 Anthropic
31.4%
as reported 		— $10 / $50 1M Vendor-reported Anthropic
14 GPT-5.6 Sol OpenAI
28.8%
max effort, v1.0.6 		$0.67 $4 / $20 1.05M Vendor-reported OpenAI
15 Qwen3.8 Max Alibaba (Qwen)
27.3%
pass@1 		— $2 / $6 1M Vendor-reported Alibaba (Qwen)
16 Claude Opus 5 Anthropic
26.9%
from Zapier's public leaderboard 		— $5 / $25 1M Vendor-reported Anthropic
17 GLM-5.2 Zhipu (Z.ai / GLM)
26.2%
v1.0.6 		— $1.4 / $4.4 1M Vendor-reported Z.ai (Zhipu)
18 GPT-6 Luna OpenAI
20.7%
max effort, v1.0.6 		$0.037 $0.1 / $0.5 1.05M Vendor-reported OpenAI
19 GPT-5.6 Luna OpenAI
17%
max effort, v1.0.6 		$0.070 $0.2 / $1.2 1.05M Vendor-reported OpenAI
20 Qwen3.7 Max Alibaba (Qwen)
14.2%
pass@1 		— $2.5 / $7.5 1M Vendor-reported Alibaba (Qwen)
§ 02 / OTHER SETTINGS

AutomationBench under other settings

Labs report AutomationBench under different settings, and a score from one setting is not comparable with a score from another. Each setting has its own table, ranked on its own.

02.1

Private set

2 entries

Google reports these on the private held-out set.

# Model Score Price / 1M Context Source
1 Gemini 3.7 Flash Deprecated Google DeepMind
30.4%
private set 		$0.75 / $3.75 1M Vendor-reported Google DeepMind
2 Gemini 3.6 Flash Google
17%
private set 		$0.75 / $3.75 1M Vendor-reported Google DeepMind
02.2

Public 600-task set

3 entries

The public set is released for experiments; leaderboard numbers come from the private set, so these scores are not comparable with the main table.

# Model Score Price / 1M Context Source
1 DeepSeek V4 Pro DeepSeek · listed as “DeepSeek V4 Pro (0813)”
31.8%
public set 		$1.32 / $3.96 1M Vendor-reported DeepSeek
2 Kimi K3 Moonshot (Kimi)
30.8%
max effort 		$3 / $15 1M Vendor-reported Moonshot AI
3 DeepSeek V4 Flash Deprecated DeepSeek · listed as “DeepSeek V4 Flash (0731)”
25.1%
public set 		$0.44 / $1.32 1M Vendor-reported DeepSeek
§ 03 / SCORE VS PRICE

AutomationBench: score against price

10% 20% 30% 40% 50% $0.030 $0.10 $0.30 $1.00 Cost per task, as published (log scale) Score GPT-5.6 Luna: 17% at $0.070GPT-5.6 Sol: 28.8% at $0.67 GPT-6 Luna: 20.7% at $0.037 GPT-6 Luna GPT-6 Sol: 33.2% at $0.27 GPT-6 Sol GPT-6.1 Sol: 36.1% at $0.30 GPT-6.1 Sol GPT-6 Astra: 41.4% at $1.73 GPT-6 Astra
The dashed line joins the models that no cheaper model beats: GPT-6 Luna (20.7%, $0.037 per task), GPT-6 Sol (33.2%, $0.27 per task), GPT-6.1 Sol (36.1%, $0.30 per task) and GPT-6 Astra (41.4%, $1.73 per task). Anything below the line costs more for a lower score. Price is the cost per task the source published for the run behind each score.
§ 04 / REASONING EFFORT

What more reasoning buys

Some publishers ran the same model at several reasoning-effort levels and printed the score and the cost per task for each. Higher effort spends more tokens per task; this is what it bought on AutomationBench.

ModelScore by effort level ($ per task)Source
GPT-6 Astra
low 30.3% $1.08 medium 34.1% $1.27 high 37.1% $1.44 xhigh 39% $1.50 max 41.4% $1.73
OpenAI
GPT-6.1 Sol
low 24.7% $0.16 medium 31.7% $0.19 high 33.2% $0.23 xhigh 35.5% $0.25 max 36.1% $0.30
OpenAI
GPT-6 Sol
low 21.2% $0.19 medium 26.9% $0.21 high 31.2% $0.24 xhigh 33.2% $0.27 max 32% $0.34
OpenAI
GPT-5.6 Sol
low 11.7% $0.31 medium 19.6% $0.42 high 24.8% $0.47 xhigh 26.3% $0.54 max 28.8% $0.67
OpenAI
GPT-6 Luna
low 1.2% $0.0060 medium 9.4% $0.016 high 14.5% $0.021 xhigh 12.6% $0.025 max 20.7% $0.037
OpenAI
GPT-5.6 Luna
low 1.8% $0.010 medium 4.3% $0.020 high 9.1% $0.050 xhigh 12.9% $0.060 max 17% $0.070
OpenAI
§ 05 / PROVENANCE

Can you trust AutomationBench scores?

No independent leaderboard publishes AutomationBench results for these models yet, so every number on this page is the lab's own report of its own model. Labs choose the harness, the effort level and the subset they report, and they rarely publish failures. Read the setting under each score before comparing two rows.

§ 06 / HOW IT WORKS

How AutomationBench works

The task

The agent lands in a simulated company with a CRM, inbox, calendar and messaging tools, gets one trigger message, and has to finish a cross-app workflow in sales, marketing, operations, support, finance or HR while ignoring irrelevant and misleading records.

The score

Pass or fail on the final state of the environment: a task counts only when the right data ended up in every system. There is no model judge; partial credit is a diagnostic, not the headline.

Test set

600 public tasks, plus a private held-out set for the leaderboard, as published by the authors.

Made by

Zapier, first published April 2026.

Official sources

§ 07 / SOURCES

Sources and freshness

26 published scores from 13 sources. Each was read on its source page; the oldest reading on this page is from November 8, 2026. Prices beside each score come from our model price index and move when a vendor reprices.

AutomationBench FAQ

Answers derived from the 26 published scores on this page and the current price index.

Q · 01

Which AI model has the highest AutomationBench score?

+
MiniMax M3.1 Flash with 62.8% (v1.0.6, 600 public tasks), as reported by MiniMax on October 19, 2026. Next: DeepSeek V4.1 Flash at 60.5% and MiMo-V2.6-Pro at 53.1%.
Q · 02

What is the cheapest model near the top of AutomationBench?

+
The leader itself. No priced model within 5 points of MiniMax M3.1 Flash is cheaper: it runs at $0.0842 per 1M tokens on our effective blend ($0.15 input, $0.6 output).
Q · 03

Are these AutomationBench scores independently verified?

+
Not yet. Every score here is the model maker's own report. Each row names the lab and links to the page where it published the number.
Q · 04

What does AutomationBench measure?

+
AutomationBench tests whether an AI agent can carry out a business workflow across several apps through their APIs, finding the endpoints itself and following company rules. The agent lands in a simulated company with a CRM, inbox, calendar and messaging tools, gets one trigger message, and has to finish a cross-app workflow in sales, marketing, operations, support, finance or HR while ignoring irrelevant and misleading records.
Q · 05

How is the AutomationBench score calculated?

+
Pass or fail on the final state of the environment: a task counts only when the right data ended up in every system. There is no model judge; partial credit is a diagnostic, not the headline. The test set has 600 public tasks, plus a private held-out set for the leaderboard.
Q · 06

How current is this AutomationBench leaderboard?

+
The newest score was published on October 19, 2026. Every row shows the date its source published it, and we last read all of them on November 8, 2026. Prices come from our model price index, where each rate carries the date it was read on the vendor's page.

One weekly digest. Zero noise.

One weekly digest · No spam, ever · Unsubscribe in one click