OSWorld

🏆 Leaderboard

As of August 20, 2026, Qwen3.8 Max is #1 for OSWorld at 86.1%. Ranked by OSWorld: can the model drive a real desktop (click, type, finish the task). OSWorld leaderboard for desktop computer-use agents, including OSWorld 2.0 and Verified, plus API pricing.

Updated August 20, 2026282 models33 providers


Qwen3.8 MaxOSS Qwen · Open Source	86.1%	86.1%	84.5%	—	$8.00
Claude Mythos 5 Anthropic · Proprietary	85%	85%	—	—	$60.00
Claude Fable 5 Anthropic · Proprietary	85%	85%	—	—	$60.00
Kimi K3OSS Moonshot AI · Open Source	84.8%	—	—	—	$18.00
Qwen3.8-27BOSS Qwen · Open Source	84.3%	84.3%	—	—	$3.65
Gemini 3.6 Flash Google · Proprietary	83%	83%	—	—	$4.50
Claude Sonnet 5 Anthropic · Proprietary	81.2%	81.2%	—	—	$12.00
GPT-5.5 OpenAI · Proprietary	78.7%	78.7%	—	13%	$35.00
Claude Sonnet 4.6 Anthropic · Proprietary	78.5%	72.1%	—	9.3%	$18.00
Seed 2.1 Turbo ByteDance · Proprietary	76.4%	—	—	—	$3.00
Gemini 3.5 Flash-Lite Google · Proprietary	74%	74%	—	—	$2.80
Claude Opus 4.6 Anthropic · Proprietary	72.7%	—	—	—	$30.00
GPT-5.4 mini OpenAI · Proprietary	72.1%	72.1%	—	—	$5.25
Claude Opus 5 Anthropic · Proprietary	70.6%	83.4%	—	70.6%	$30.00
Qwen3 VL 235B A22B InstructOSS Qwen · Open Source	66.7%	—	62%	—	$1.79
Claude Opus 4.5 Anthropic · Proprietary	66.3%	—	—	—	$30.00
GPT-5.4 OpenAI · Proprietary	64.9%	75%	—	—	$17.50
Kimi K2.5OSS Moonshot AI · Open Source	63.3%	63.3%	—	—	$3.68
GLM-5V-TurboOSS Z AI · Open Source	62.3%	—	—	—	$5.20
Claude Sonnet 4.5 Anthropic · Proprietary	61.4%	62.9%	—	—	$18.00
GPT-5.3 Codex OpenAI · Proprietary	59%	64.7%	—	—	$15.75
Claude Haiku 4.5 Anthropic · Proprietary	50.7%	—	—	—	$6.00
GPT-5.2 Pro OpenAI · Proprietary	47.8%	—	—	—	$189.00
Gemini 3 Pro Google · Proprietary	47.5%	47.5%	72.7%	—	$14.00
GPT-5.2 OpenAI · Proprietary	39.2%	—	86.3%	—	$15.75
GPT-5.4 nano OpenAI · Proprietary	39%	39%	—	—	$1.45
Claude Sonnet 4 Anthropic · Proprietary	38.6%	—	—	—	$18.00
Qwen3 VL 235B A22B ThinkingOSS Qwen · Open Source	38.1%	—	61.8%	—	$3.94
Claude 3.7 Sonnet Anthropic · Proprietary	35.8%	35.8%	—	—	$18.00
Qwen3 VL 8B ThinkingOSS Qwen · Open Source	33.9%	—	46.6%	—	$2.27
Qwen3 VL 8B InstructOSS Qwen · Open Source	33.9%	—	54.6%	—	$0.58
Qwen3 VL 32B InstructOSS Qwen · Open Source · via OpenRouter	32.6%	—	57.9%	—	$0.52
Qwen3 VL 4B ThinkingOSS Qwen · Open Source	31.4%	—	49.2%	—	$1.10
Qwen3 VL 30B A3B ThinkingOSS Qwen · Open Source	30.6%	—	57.3%	—	$1.20
Qwen3 VL 30B A3B InstructOSS Qwen · Open Source	30.3%	—	60.5%	—	$0.90
Qwen3 VL 4B InstructOSS Qwen · Open Source	26.2%	—	59.5%	—	$0.70
o3 OpenAI · Proprietary	23%	23%	—	—	$10.00
Qwen2.5 VL 72B InstructOSS Qwen · Open Source · via OpenRouter	8.8%	5%	43.6%	—	$1.80
ChatGPT-4o Latest OpenAI · Proprietary	—	—	—	—	$12.50
Claude 3 Haiku Anthropic · Proprietary	—	—	—	—	$1.50
Claude 3 Opus Anthropic · Proprietary	—	—	—	—	$90.00
Claude 3 Sonnet Anthropic · Proprietary	—	—	—	—	$18.00
Claude 3.5 Haiku Anthropic · Proprietary	—	—	—	—	$4.80
Claude 3.5 Sonnet Anthropic · Proprietary	—	—	—	—	$18.00
Claude 3.5 Sonnet Anthropic · Proprietary	—	—	—	—	$18.00
Claude Mythos Preview Anthropic · Proprietary	—	79.6%	—	—	$60.00
Claude Opus 4 Anthropic · Proprietary	—	—	—	—	$90.00
Claude Opus 4.1 Anthropic · Proprietary	—	—	—	—	$90.00
Claude Opus 4.7 Anthropic · Proprietary	—	78%	—	18.2%	$30.00
Claude Opus 4.8 Anthropic · Proprietary	—	83.4%	87.9%	20.6%	$30.00

Showing 1–50 of 282 models

Next step

You found the model. Now ship the product.

Auth, billing, and the API layer are already decided. Fork 8 finished AI apps, or have us build the first version with you.

Get the stack — $249 Have us build it with you

The index

All Large Language Models

282 models across 33 providers. Search or jump to a lab — every model page stays linked here.

Cursor

2 models

Composer 2 Composer 2 Fast

xAI

15 models

Grok 4.6 Grok 4.5 Grok 4.3 Grok-4.1 Grok-4.1 Fast Non-Reasoning Grok-4.1 Fast Reasoning Grok-4.1 Thinking Grok 4 Fast Grok Code Fast 1 Grok-4 Fast Non-Reasoning Grok-4 Fast Reasoning Grok-4 Grok-3 Grok-3 Mini Grok-2

DeepSeek

16 models

DeepSeek-V4-Pro-0813 DeepSeek-V4-Flash-0731 DeepSeek-V4-Flash-0423 DeepSeek-V4-Flash-Max DeepSeek-V4-Pro-Max DeepSeek-V3.2 DeepSeek-V3.2 (Non-thinking)DeepSeek-V3.2-Exp DeepSeek-R1-0528 DeepSeek-V3 0324 DeepSeek R1 Distill Llama 70B DeepSeek R1 Distill Qwen 32B DeepSeek-R1 DeepSeek-V3.1 DeepSeek-V3 DeepSeek-V2.5

MiniMax

6 models

MiniMax M3 MiniMax M2.7 MiniMax M2.5 MiniMax M2.1 MiniMax M2 MiniMax M1 80K

AI21 Labs

2 models

Jamba 1.5 Large Jamba 1.5 Mini

Amazon

5 models

Nova 2 Lite Nova 2 Sonic Nova Lite Nova Micro Nova Pro

Baidu

1 models

ERNIE 4.5

ByteDance

2 models

Seed 2.1 Turbo Seed 2.0 Pro

Cohere

3 models

North Mini Code 1.0 Command A+Command R+

IBM

1 models

Granite 3.3 8B Instruct

Inception

1 models

Mercury 2

inclusionAI

1 models

Ling-3.0-flash

LG AI Research

1 models

K-EXAONE-236B-A23B

Meituan

4 models

LongCat-Flash-Lite LongCat-Flash-Thinking-2601 LongCat-Flash-Thinking LongCat-Flash-Chat

Microsoft

5 models

MAI-Code-1.1-Flash Phi 4 Mini Phi-4-multimodal-instruct Phi 4 Phi-3.5-mini-instruct

Moonshot AI

9 models

Kimi K3 Kimi K2.7 Code Kimi K2.7 Code HighSpeed Kimi K2.6 Kimi K2.5 Kimi K2 0905 Kimi K2-Instruct-0905 Kimi K2-Thinking-0905 Kimi K2 Instruct

Nous Research

1 models

Hermes 3 70B

NVIDIA

5 models

Nemotron 3.5 Lightning (30B A3B)Nemotron 3 Ultra (550B A55B)Nemotron 3 Super (120B A12B)Nemotron 3 Nano (30B A3B)Nemotron Nano 9B v2

Poolside

2 models

Laguna S 2.1 Laguna XS 2.1

Sakana AI

1 models

Sakana Namazu

StepFun

1 models

Step-3.5-Flash

Tencent

1 models

Hy3

Thinking Machines

1 models

Inkling-Small

Unisound

1 models

Upstage

1 models

Solar Pro 4

Xiaomi

3 models

MiMo-V2.5-Pro MiMo-V2.5 MiMo-V2-Flash

Which model leads OSWorld right now?

As of August 20, 2026, Qwen3.8 Max by Qwen is #1 for OSWorld at 86.1%. Ranked by OSWorld: can the model drive a real desktop (click, type, finish the task). This board also tracks OSWorld, OSWorld Verified, ScreenSpot Pro, OSWorld 2.0. Next on the same board: Claude Fable 5 and Claude Mythos 5. Related leaders: Claude Opus 4.8 on ScreenSpot Pro at 87.9%; Claude Opus 5 on OSWorld 2.0 at 70.6%. This osworld leaderboard ranks models by OSWorld. Scores come from public evals. Prices are the live API rates in the table above.

Sources: OSWorld (https://os-world.github.io/); OpenAI API pricing (https://developers.openai.com/api/docs/pricing); Anthropic Claude API pricing (https://platform.claude.com/docs/en/about-claude/pricing); Gemini API pricing (https://ai.google.dev/gemini-api/docs/pricing)

Top 8 for osworld. Ranked by OSWorld: can the model drive a real desktop (click, type, finish the task). Input and output are dollars per million tokens.
Rank	Model	OSWorld	Input /M	Output /M
1	Qwen3.8 Max	86.1%	$2.00	$6.00
2	Claude Fable 5	85%	$10.00	$50.00
3	Claude Mythos 5	85%	$10.00	$50.00
4	Kimi K3	84.8%	$3.00	$15.00
5	Qwen3.8-27B	84.3%	$0.45	$3.20
6	Gemini 3.6 Flash	83%	$0.75	$3.75
7	Claude Sonnet 5	81.2%	$2.00	$10.00
8	GPT-5.5	78.7%	$5.00	$30.00

Related LLM leaderboards

LLM Leaderboard Coding Math Agentic Reasoning Writing Science Finance Long Context Vision Open Source

Individual benchmark pages

Rank one eval at a time. All LLM benchmarks.

HLE SWE Bench Humanity's Last Exam (No Tools)Humanity's Last Exam (With Tools)Terminal-Bench 2.0 Terminal-Bench Terminal-Bench 2.1 Terminal-Bench 2 Terminal-Bench 3.0 ARC-AGI-1 Verified

OSWorld FAQ

Who ranks #1 on the OSWorld leaderboard?

As of August 20, 2026, Qwen3.8 Max by Qwen ranks #1 on OSWorld at 86.1%. API pricing is $2.00/M input and $6.00/M output.

What are the top models on OSWorld?

The current OSWorld ranking as of August 20, 2026 is 1. Qwen3.8 Max at 86.1%; 2. Claude Fable 5 at 85%; 3. Claude Mythos 5 at 85%.

Which osworld model is the cheapest?

Qwen3 VL 32B Instruct is the cheapest scored model on this osworld leaderboard at $0.10/M input and $0.42/M output ($0.52 blended). Qwen3.8 Max still leads OSWorld at 86.1%.

Should I always pick the #1 OSWorld model?

Not automatically. Qwen3.8 Max leads OSWorld, but a cheaper scored model can be the better production choice if the quality gap is small. Use the table to weigh OSWorld against input/output price, context window, and related evals.

How often is the OSWorld leaderboard updated?

Scores and API prices on this page are refreshed from published evals and provider rates. The snapshot is labeled August 20, 2026. Treat it as a current index, not a one-off blog post.

What is OSWorld?

OSWorld is a computer-use benchmark. The agent has to operate a real desktop: open apps, click UI, and finish a task end to end.

What is OSWorld Verified?

OSWorld Verified is a cleaned task subset. OSWorld 2.0 is the newer harder suite. We show every published variant.

Is OSWorld the same as Terminal-Bench?

No. Terminal-Bench is shell-only. OSWorld includes GUI grounding, which is why ScreenSpot Pro sits on this page as a related column.

You found the model. Now ship the product.

All Large Language Models

OpenAI

Anthropic

Google

Cursor

xAI

DeepSeek

MiniMax

Qwen

AI21 Labs

Amazon

Baidu

ByteDance

Cohere

IBM

Inception

inclusionAI

LG AI Research

Meituan

Meta

Microsoft

Mistral

Moonshot AI

Nous Research

NVIDIA

Poolside

Sakana AI

StepFun

Tencent

Thinking Machines

Unisound

Upstage

Xiaomi

Z AI

Which model leads OSWorld right now?

Related LLM leaderboards

Individual benchmark pages

OSWorld FAQ

Who ranks #1 on the OSWorld leaderboard?

What are the top models on OSWorld?

Which osworld model is the cheapest?

Should I always pick the #1 OSWorld model?

How often is the OSWorld leaderboard updated?

What is OSWorld?

What is OSWorld Verified?

Is OSWorld the same as Terminal-Bench?