Last verified 2026-11-08

Agents and computer useVendor-reportedPrices verified

OSWorld 2 leaderboard

As of November 8, 2026, Claude Opus 5.5 (Anthropic) leads OSWorld 2 at 81.8%, ahead of Claude Fable 5.1 at 80.7% and Claude Sonnet 5.5 at 80.1%. 19 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

81.8%

Claude Opus 5.5

Systems ranked

across 5 settings

Independent rows

0/20

all vendor-reported

Cheapest near top

$1.28/1M

Kimi K3.1 Pro · within 5 pts

With API prices

19/19

rates from our price index

Newest score

Nov 2

2026 · newest result

What is OSWorld 2? OSWorld 2 tests computer-use agents on long, realistic desktop and web workflows that take a person about an hour and a half and an agent hundreds of actions. Made by XLANG Lab at the University of Hong Kong with university and industry partners, June 2026 (2.1 bug-fix release in September 2026).

§ 01 / LEADERBOARD

Top models on OSWorld 2, November 2026

6 entries · 2.1, partial score

#	Model	Score	Price / 1M	Context	Source
1	Claude Opus 5.5 Anthropic	81.8% partial reward, max effort	$4 / $20	1M	Vendor-reported Anthropic 2026-09-22
2	Claude Fable 5.1 Anthropic	80.7% partial reward, max effort	$10 / $50	1M	Vendor-reported Anthropic 2026-09-22
3	Claude Sonnet 5.5 Anthropic	80.1% partial reward	$2 / $10	1M	Vendor-reported Anthropic 2026-09-28
4	Kimi K3.1 Pro Moonshot (Kimi)	79.6% max effort	$2 / $10	1M	Vendor-reported Moonshot AI 2026-11-02
5	Claude Opus 5 Anthropic	74% partial reward, max effort	$5 / $25	1M	Vendor-reported Anthropic 2026-09-22
6	Claude Sonnet 5 Anthropic	57% partial reward	$2 / $10	1M	Vendor-reported Anthropic 2026-09-28

§ 02 / OTHER SETTINGS

OSWorld 2 under other settings

Labs report OSWorld 2 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

2.0, offline subset, partial score

7 entries

#	Model	Score	$ / task	Price / 1M	Context	Source
1	GPT-6 Astra OpenAI	73.5% max effort	$9.07	$10 / $50	1.05M	Vendor-reported OpenAI 2026-09-22
2	GPT-6.1 Sol OpenAI	71.4% max effort	$1.27	$2 / $10	1.05M	Vendor-reported OpenAI 2026-09-29
3	Gemini 4 Argon Google DeepMind	69.2% offline subset, partial score	—	$2 / $10	1M	Vendor-reported Google DeepMind 2026-10-13
4	GPT-5.6 Sol OpenAI	66.2% max effort	$7.71	$4 / $20	1.05M	Vendor-reported OpenAI 2026-09-22
5	GPT-6 Sol OpenAI	64.4% max effort	$3.25	$2 / $10	1.05M	Vendor-reported OpenAI 2026-09-22
6	GPT-5.6 Luna OpenAI	52.7% max effort	$0.49	$0.2 / $1.2	1.05M	Vendor-reported OpenAI 2026-09-22
6	GPT-6 Luna OpenAI	52.7% max effort	$0.27	$0.1 / $0.5	1.05M	Vendor-reported OpenAI 2026-09-22

02.2

2.1, offline subset

3 entries

#	Model	Score	Price / 1M	Context	Source
1	Claude Sonnet 5.5 Anthropic	83.9% offline subset	$2 / $10	1M	Vendor-reported Anthropic 2026-10-07
2	Claude Haiku 5.5 Anthropic	72.4% offline subset	$0.1 / $0.5	1M	Vendor-reported Anthropic 2026-10-07
3	Claude Haiku 4.5 Anthropic	15.7% offline subset	$1 / $5	200K	Vendor-reported Anthropic 2026-10-07

02.3

2.0, partial score, batch tool on

2 entries

#	Model	Score	Price / 1M	Context	Source
1	Gemini 3.8 Flash Google DeepMind	59% partial score, batch tool enabled	$0.75 / $3.75	1M	Vendor-reported Google DeepMind 2026-09-02
2	Gemini 3.7 Flash Deprecated Google DeepMind	50.6% partial score, batch tool enabled	$0.75 / $3.75	1M	Vendor-reported Google DeepMind 2026-09-02

02.4

2.0, scoring not stated

2 entries

#	Model	Score	Price / 1M	Context	Source
1	MiniMax M3.1 Flash MiniMax	61.7% scoring not stated	$0.15 / $0.6	1M	Vendor-reported MiniMax 2026-10-19
2	Kimi K3 Moonshot (Kimi)	58.3% max effort, scoring not stated	$3 / $15	1M	Vendor-reported Moonshot AI 2026-07-16

§ 03 / SCORE VS PRICE

OSWorld 2: score against price

The dashed line joins the models that no cheaper model beats: Kimi K3.1 Pro (79.6%, $1.28 per 1M tokens), Claude Sonnet 5.5 (80.1%, $1.29 per 1M tokens) and Claude Opus 5.5 (81.8%, $2.58 per 1M tokens). Anything below the line costs more for a lower score. Price is the effective rate per 1M tokens on current list prices: 92% input, 8% output, 82% of input served from cache.

§ 04 / PROVENANCE

Can you trust OSWorld 2 scores?

No independent leaderboard publishes OSWorld 2 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.

!Labs report different releases (2.0, 2.1), scoring (partial or full completion) and subsets, and the gap between partial and full scores is large. Each combination has its own table.
!Some labs report an offline subset of tasks; the official pages we read do not define it.
!Step budget and effort setting change the score; the official headline uses 500 steps.

§ 05 / HOW IT WORKS

How OSWorld 2 works

The task

Each task is an end-to-end workflow, for example filing an expense claim across email, a bank site and an internal portal, often with hidden state or new information arriving midway.

The score

The main metric is whether the whole task was completed within 500 steps; a partial score averages about 27 checkpoints per task. Labs quote either, so the two are separate tables here.

Test set

108 tasks, as published by the authors.

Made by

XLANG Lab at the University of Hong Kong with university and industry partners, first published June 2026 (2.1 bug-fix release in September 2026).

Official sources

§ 06 / SOURCES

Sources and freshness

20 published scores from 10 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.

Vendor-reported OpenAI 5 scores · newest 2026-09-22
Vendor-reported Anthropic 3 scores · newest 2026-09-22
Vendor-reported Anthropic 3 scores · newest 2026-10-07
Vendor-reported Anthropic 2 scores · newest 2026-09-28
Vendor-reported Google DeepMind 2 scores · newest 2026-09-02
Vendor-reported Google DeepMind 1 score · newest 2026-10-13
Vendor-reported MiniMax 1 score · newest 2026-10-19
Vendor-reported Moonshot AI 1 score · newest 2026-11-02
Vendor-reported Moonshot AI 1 score · newest 2026-07-16
Vendor-reported OpenAI 1 score · newest 2026-09-29

OSWorld 2 FAQ

Answers derived from the 20 published scores on this page and the current price index.

Q · 01

Which AI model has the highest OSWorld 2 score?

Claude Opus 5.5 with 81.8% (partial reward, max effort), as reported by Anthropic on September 22, 2026. Next: Claude Fable 5.1 at 80.7% and Claude Sonnet 5.5 at 80.1%.

Q · 02

What is the cheapest model near the top of OSWorld 2?

Kimi K3.1 Pro scores 79.6%, 2.2 points behind the leader, at $1.28 per 1M tokens on our effective blend ($2 input, $10 output). Claude Opus 5.5 costs $2.58 on the same blend.

Q · 03

Are these OSWorld 2 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 OSWorld 2 measure?

OSWorld 2 tests computer-use agents on long, realistic desktop and web workflows that take a person about an hour and a half and an agent hundreds of actions. Each task is an end-to-end workflow, for example filing an expense claim across email, a bank site and an internal portal, often with hidden state or new information arriving midway.

Q · 05

How is the OSWorld 2 score calculated?

The main metric is whether the whole task was completed within 500 steps; a partial score averages about 27 checkpoints per task. Labs quote either, so the two are separate tables here. The test set has 108 tasks.

Q · 06

How current is this OSWorld 2 leaderboard?

The newest score was published on November 2, 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.