OSWorld 2.1(offline subset)

Offline subset of OSWorld 2.1 (as labeled on the Claude Haiku 5.5 announcement page).

Category: Computer use
Unit: Percent (%); Higher is better
Comparison groups: 1total; Currently available model rankings · Ranked results: 15; All models (records): Groups: 1 · Entries: 16

Rankings

Values with different source types (official, third-party, paper) or measurement conditions cannot be compared, so rankings are separated by comparison group (the same source, or sources verified to use the same conditions).

Rankings include only models currently available in Claude Code / Codex, excluding legacy models.

official results (Introducing Claude Haiku 5.5)

Rank	Model	effort	Score	Source cost (metric)	measured/reported by	source
1	Claude Sonnet 5.5	Max	83.9%	$5.73Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Sonnet 5.5 · Max: 83.9% $5.73"
2	Claude Sonnet 5.5	Xhigh	81.1%	$2.22Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Sonnet 5.5 · Xhigh: 81.1% $2.22"
3	Claude Sonnet 5.5	High	73.2%	$1.38Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Sonnet 5.5 · High: 73.2% $1.38"
4	Claude Haiku 5.5	Max	72.4%	$0.61Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Haiku 5.5 · Max: 72.4% $0.61"
5	Claude Haiku 5.5	Xhigh	67.6%	$0.28Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Haiku 5.5 · Xhigh: 67.6% $0.28"
6	Claude Sonnet 5.5	Med	66.0%	$0.93Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Sonnet 5.5 · Med: 66.0% $0.93"
7	Claude Haiku 5.5	High	61.3%	$0.18Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Haiku 5.5 · High: 61.3% $0.18"
8	Claude Sonnet 5.5	Low	57.9%	$0.68Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Sonnet 5.5 · Low: 57.9% $0.68"
9	Claude Haiku 5.5	Med	53.3%	$0.13Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Haiku 5.5 · Med: 53.3% $0.13"
10	GPT-6 Luna	Max	48.9%	$0.20Cost per attempt	not stated	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "Max" was mapped to max in the OpenAI API reasoning effort Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "GPT-6 Luna · Max: 48.9% $0.20"
11	GPT-6 Luna	Xhigh	44.8%	$0.17Cost per attempt	not stated	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "Xhigh" was mapped to xhigh in the OpenAI API reasoning effort Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "GPT-6 Luna · Xhigh: 44.8% $0.17"
12	GPT-6 Luna	High	42.3%	$0.14Cost per attempt	not stated	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "High" was mapped to high in the OpenAI API reasoning effort Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "GPT-6 Luna · High: 42.3% $0.14"
13	Claude Haiku 5.5	Low	42.0%	$0.07Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Haiku 5.5 · Low: 42.0% $0.07"
14	GPT-6 Luna	Med	37.5%	$0.13Cost per attempt	not stated	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "Med" was mapped to medium in the OpenAI API reasoning effort Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "GPT-6 Luna · Med: 37.5% $0.13"
15	GPT-6 Luna	Low	19.2%	$0.04Cost per attempt	not stated	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "Low" was mapped to low in the OpenAI API reasoning effort Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "GPT-6 Luna · Low: 19.2% $0.04"

All model rankings (records)

These rankings are recalculated by comparison group including legacy models and models unavailable in Claude Code / Codex. They differ from the rankings above because the set of models differs.

Show rankings including legacy and unavailable models (entries: 16)

official results (Introducing Claude Haiku 5.5)

Rank	Model	effort	Score	Source cost (metric)	measured/reported by	source
1	Claude Sonnet 5.5	Max	83.9%	$5.73Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Sonnet 5.5 · Max: 83.9% $5.73"
2	Claude Sonnet 5.5	Xhigh	81.1%	$2.22Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Sonnet 5.5 · Xhigh: 81.1% $2.22"
3	Claude Sonnet 5.5	High	73.2%	$1.38Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Sonnet 5.5 · High: 73.2% $1.38"
4	Claude Haiku 5.5	Max	72.4%	$0.61Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Haiku 5.5 · Max: 72.4% $0.61"
5	Claude Haiku 5.5	Xhigh	67.6%	$0.28Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Haiku 5.5 · Xhigh: 67.6% $0.28"
6	Claude Sonnet 5.5	Med	66.0%	$0.93Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Sonnet 5.5 · Med: 66.0% $0.93"
7	Claude Haiku 5.5	High	61.3%	$0.18Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Haiku 5.5 · High: 61.3% $0.18"
8	Claude Sonnet 5.5	Low	57.9%	$0.68Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Sonnet 5.5 · Low: 57.9% $0.68"
9	Claude Haiku 5.5	Med	53.3%	$0.13Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Haiku 5.5 · Med: 53.3% $0.13"
10	GPT-6 Luna	Max	48.9%	$0.20Cost per attempt	not stated	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "Max" was mapped to max in the OpenAI API reasoning effort Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "GPT-6 Luna · Max: 48.9% $0.20"
11	GPT-6 Luna	Xhigh	44.8%	$0.17Cost per attempt	not stated	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "Xhigh" was mapped to xhigh in the OpenAI API reasoning effort Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "GPT-6 Luna · Xhigh: 44.8% $0.17"
12	GPT-6 Luna	High	42.3%	$0.14Cost per attempt	not stated	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "High" was mapped to high in the OpenAI API reasoning effort Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "GPT-6 Luna · High: 42.3% $0.14"
13	Claude Haiku 5.5	Low	42.0%	$0.07Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic) Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Haiku 5.5 · Low: 42.0% $0.07"
14	GPT-6 Luna	Med	37.5%	$0.13Cost per attempt	not stated	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "Med" was mapped to medium in the OpenAI API reasoning effort Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "GPT-6 Luna · Med: 37.5% $0.13"
15	GPT-6 Luna	Low	19.2%	$0.04Cost per attempt	not stated	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "Low" was mapped to low in the OpenAI API reasoning effort Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "GPT-6 Luna · Low: 19.2% $0.04"
16	Claude Haiku 4.5	not stated	15.7%	$1.45Cost per attempt	Anthropic	Introducing Claude Haiku 5.5Retrieved 2026-11-12 · Vendor-reported value. A point in the figure "OSWorld 2.1 (offline subset) / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for its own model (announced by Anthropic); no effort because the source effort label "max" is not in the official effort list Source excerpt OSWorld 2.1 (offline subset) — Accuracy vs. cost: "Haiku 4.5 · Max: 15.7% $1.45"