OSWorld Leaderboard 2026: Latest Computer Use Agent Scores

Scores

Open weightsClosed weightsMixed / uncertain

How model sizes and weights are classified

Size classes follow the Artificial Analysis categories: Tiny ≤4B; Small >4–40B; Medium >40–150B; Large >150B total parameters. Active MoE parameters are reported separately.

When parameters are unavailable, we use standard uncached token prices with a 3:1 input/output blend: Small <$0.15/M, Medium $0.15–$1/M, Large >$1/M. “Price proxy” is a pricing category, not a parameter estimate. “Estimated” marks a provisional classification and remains included in filters. “Unclassified” results await model research and match only the all and mixed / uncertain weight filters.

Harnesses inherit the evaluated model stack's class. Multi-model systems use the largest component category as an approximation. Open weights requires released weights for every required model; open harness code alone does not qualify. Unverified backbones and unavailable evaluated checkpoints are marked uncertain.

Results compare evaluated systems, including their tools, compute budgets, and context strategies. Size categories do not guarantee hardware compatibility. Suggest a correction ↗

Results

Agent scope · 23 systems

Rank	System / submission	Score ↓	Model class	Reported	Source
1	Qwen3.8-MaxNewAlibaba · Model Evaluation & model details OSWorld-Verified from the official launch eval table, which also lists Opus 4.8, Fable 5, Gemini 3.1 Pro, GPT-5.6 Sol, and Qwen3.7-Plus baselines. Self-reported by Alibaba. Preserves the launch evaluation of the visual Max model. Related text-only open weights are insufficient to mark this GUI result open. Harness code: uncertain. Published size: from the parameter count of Qwen3.8-Max (August 2026 launch). Reported OSWorld evaluation Alibaba Qwen3.8-Max launch announcement, August 3, 2026 Qwen/Qwen3.8-2.4T-A95B model card and released files Qwen3.8-Max (August 2026 launch) Uncertain weights · Large · Published size · 2,400B total Alibaba publishes 2.4T total parameters for the evaluated multimodal flagship. The related public Qwen3.8-2.4T-A95B checkpoint has 213 weight shards, but its card explicitly says text-only and distinguishes the hosted Max version with vision. That release does not establish availability of the evaluated visual checkpoint. The live Max alias was upgraded in September; this record preserves the August launch identity. Alibaba Qwen3.8-Max launch announcement, August 3, 2026 Qwen/Qwen3.8-2.4T-A95B model card and released files Qwen/Qwen3.8-2.4T-A95B checkpoint inventory Qwen3.8-Max September alias upgrade notice Checked 2026-10-10 · medium confidence	86.1%	Uncertain weightsLarge	2026-08-03	Source ↗
2	Claude Mythos PreviewNewAnthropic · Model Evaluation & model details OSWorld-Verified pass@1 (361 tasks, 100 max steps, auto-thinking at max effort, 5-run avg) on Anthropic's revised harness; supersedes the earlier 79.6% Glasswing report. Self-reported in the Claude 5 system card. Reuses the exact Claude model version already researched for BrowseComp. The retrieved Claude 5 system card section 8.16.3 describes a revised OSWorld computer-use harness, 100 steps and five runs; no separate inference model is identified. Full internal evaluation harness availability is unverified. Harness code: uncertain. Price proxy: the size class comes from the standard token price of Claude Mythos Preview; parameters are undisclosed. Reported OSWorld evaluation Claude Mythos Preview Closed weights · Large · Price proxy Token price: $25 input / $125 output per million. Blended 3:1: $50/M. Undisclosed parameters. Published participant API price; limited research-preview access, with initial usage credits. Token pricing excludes tools. Anthropic Project Glasswing Checked 2026-10-10 · high confidence	85.4%	Closed weightsLargePrice proxy	2026-06	Source ↗
3	Claude Mythos 5NewAnthropic · Model Evaluation & model details OSWorld-Verified pass@1 (361 tasks, 100 max steps, auto-thinking at max effort, 5-run avg) on Anthropic's revised harness. Self-reported in the Claude 5 system card. Reuses the exact Claude model version already researched for BrowseComp. The retrieved Claude 5 system card section 8.16.3 describes a revised OSWorld computer-use harness, 100 steps and five runs; no separate inference model is identified. Full internal evaluation harness availability is unverified. Harness code: uncertain. Price proxy: the size class comes from the standard token price of Claude Mythos 5; parameters are undisclosed. Reported OSWorld evaluation Claude Mythos 5 Closed weights · Large · Price proxy Token price: $10 input / $50 output per million. Blended 3:1: $20/M. Undisclosed parameters; size uses standard uncached token pricing. Tool charges excluded. Anthropic model pricing Checked 2026-10-10 · high confidence	85.0%	Closed weightsLargePrice proxy	2026-06	Source ↗
4	Claude Fable 5NewAnthropic · Model Evaluation & model details Generally available sibling of Mythos 5; OSWorld-Verified pass@1 (361 tasks, 100 max steps, auto-thinking at max effort, 5-run avg). Self-reported in the Claude 5 system card. Reuses the exact Claude model version already researched for BrowseComp. The retrieved Claude 5 system card section 8.16.3 describes a revised OSWorld computer-use harness, 100 steps and five runs; no separate inference model is identified. Full internal evaluation harness availability is unverified. Harness code: uncertain. Price proxy: the size class comes from the standard token price of Claude Fable 5; parameters are undisclosed. Reported OSWorld evaluation Claude Fable 5 Closed weights · Large · Price proxy Token price: $10 input / $50 output per million. Blended 3:1: $20/M. Undisclosed parameters; size uses standard uncached token pricing. Tool charges excluded. Anthropic model pricing Checked 2026-10-10 · high confidence	85.0%	Closed weightsLargePrice proxy	2026-06	Source ↗
5	Claude Opus 4.8NewAnthropic · Model Evaluation & model details OSWorld-Verified pass@1 (361 tasks, 100 steps, max effort, multi-run avg) using Anthropic's revised harness (zoom-tool fix, 128K tokens/turn) — not directly comparable to older entries. Self-reported in the Opus 4.8 system card. Named model on Anthropic’s revised computer-use harness. Reuses its existing model record; the launch explicitly notes evaluation changes, which remain in the result notes. Harness code: uncertain. Price proxy: the size class comes from the standard token price of Claude Opus 4.8; parameters are undisclosed. Reported OSWorld evaluation Claude Opus 4.8 Closed weights · Large · Price proxy Token price: $5 input / $25 output per million. Blended 3:1: $10/M. Undisclosed parameters; size uses standard uncached token pricing. Tool charges excluded. Anthropic model pricing Checked 2026-10-10 · high confidence	83.4%	Closed weightsLargePrice proxy	2026-05-28	Source ↗
6	MiMo-V2.6-ProNewXiaomi · Model Evaluation & model details OSWorld-Verified score from Table 3 of the MiMo-V2.6 technical report and the official model card. Self-reported by Xiaomi; step budget, run count, and harness settings are not stated for this result. Maps the official result to the released MiMo-V2.6-Pro-RL checkpoint; card and actual weight inventory verified. No additional runtime model is identified. Harness code: not separately classified. Published size: from the parameter count of MiMo-V2.6-Pro-RL. Reported OSWorld evaluation MiMo-V2.6-Pro-RL Open weights · Large · Published size · 1,020B total · 42B active The official MIT-licensed card identifies this released RL checkpoint and publishes 1020B total/42B activated parameters. Public inventory contains 132 safetensor files. Its stored tensor inventory is about 1024.2B including auxiliary/quantization tensors; the explicitly published architectural total drives classification. OSWorld numbers are in the shared evaluation table; no separate inference helper is identified. XiaomiMiMo/MiMo-V2.6-Pro-RL model card and released files XiaomiMiMo/MiMo-V2.6-Pro-RL checkpoint inventory Checked 2026-10-10 · high confidence	82.0%	Open weightsLarge	2026-09-22	Source ↗Repository ↗
7	Claude Sonnet 5NewAnthropic · Model Evaluation & model details OSWorld-Verified pass@1 (361 tasks, 100 max steps) on Anthropic's revised harness. Self-reported in the Claude Sonnet 5 system card. Reuses the exact Sonnet 5 model. Retrieved system card section 8.10.2 identifies the revised 100-step GUI evaluation; no separate inference helper is specified. Harness code: uncertain. Price proxy: the size class comes from the standard token price of Claude Sonnet 5; parameters are undisclosed. Reported OSWorld evaluation Claude Sonnet 5 Closed weights · Large · Price proxy Token price: $2 input / $10 output per million. Blended 3:1: $4/M. Undisclosed parameters; size uses standard uncached token pricing. Tool charges excluded. Anthropic model pricing Checked 2026-10-10 · high confidence	81.2%	Closed weightsLargePrice proxy	2026-06-30	Source ↗
8	MiMo-V2.6-FlashNewXiaomi · Model Evaluation & model details OSWorld-Verified score from Table 3 of the MiMo-V2.6 technical report and the official model card. Self-reported by Xiaomi; step budget, run count, and harness settings are not stated for this result. Maps the official result to the released MiMo-V2.6-Flash-RL checkpoint; card and actual weight inventory verified. No additional runtime model is identified. Harness code: not separately classified. Published size: from the parameter count of MiMo-V2.6-Flash-RL. Reported OSWorld evaluation MiMo-V2.6-Flash-RL Open weights · Large · Published size · 309B total · 15B active The official MIT-licensed card identifies this released RL checkpoint and publishes 309B total/15B activated parameters. Public inventory contains 67 safetensor files. Its stored tensor inventory is about 310.8B including auxiliary/quantization tensors; the explicitly published architectural total drives classification. OSWorld numbers are in the shared evaluation table; no separate inference helper is identified. XiaomiMiMo/MiMo-V2.6-Flash-RL model card and released files XiaomiMiMo/MiMo-V2.6-Flash-RL checkpoint inventory Checked 2026-10-10 · high confidence	80.8%	Open weightsLarge	2026-09-22	Source ↗Repository ↗
9	Claude Sonnet 4.6Anthropic · Model Evaluation & model details OSWorld-Verified pass@1 (361 tasks, 100 max steps) on Anthropic's revised harness. Self-reported in the Claude Sonnet 5 system card. Sonnet 4.6 was re-evaluated in the Sonnet 5 card using the revised zoom tool and per-turn token limit. It remains the Sonnet 4.6 model, with the changed harness preserved in result context. Harness code: uncertain. Price proxy: the size class comes from the standard token price of Claude Sonnet 4.6; parameters are undisclosed. Reported OSWorld evaluation Claude Sonnet 4.6 Closed weights · Large · Price proxy Token price: $3 input / $15 output per million. Blended 3:1: $6/M. Undisclosed parameters; size uses standard uncached token pricing. Tool charges excluded. Anthropic model pricing Checked 2026-10-10 · high confidence	78.5%	Closed weightsLargePrice proxy	2026-06-30	Source ↗
10	OSAgentTheAGI Company · System Evaluation & model details Self-reported October 2025; trained with RL on OSWorld VMs and internal browser environments. The launch identifies a trained policy, but not its general-reasoning base or released checkpoint. Runtime identity and full implementation remain unverified; Large is explicitly provisional. Harness code: uncertain. Estimated for OSAgent (October 2025 undisclosed policy): no verified parameter count or standard token price, so the size class is provisional. Reported OSWorld evaluation OSAgent (October 2025 undisclosed policy) Uncertain weights · Large · Estimated The launch describes an end-to-end RL computer-use policy trained from an unnamed general-reasoning base. Neither exact backbone, parameter count, token prices, nor evaluated weights are published there. Large is a provisional frontier reasoning-policy estimate, not a measured parameter claim. An agi-inc 32B repository created in April 2026 has no model card and cannot establish this October 2025 model identity. OSAgent launch and training description Later AGI checkpoint metadata; not proof of historical identity Checked 2026-10-10 · low confidence	76.26%	Uncertain weightsLargeEstimated	2025-10-23	Source ↗
11	GPT-5.4NewOpenAI · Model Evaluation & model details Self-reported at GPT-5.4 launch on OSWorld-Verified; awaiting independent verification. The launch reports native screenshot-based computer use for GPT-5.4. Reuses the GPT-5.4 record; no additional model is identified for this result. Harness code: uncertain. Price proxy: the size class comes from the standard token price of GPT-5.4; parameters are undisclosed. Reported OSWorld evaluation GPT-5.4 Closed weights · Large · Price proxy Token price: $2.5 input / $15 output per million. Blended 3:1: $5.625/M. Parameters are undisclosed. Category uses published standard uncached API token prices, not an estimated parameter count. Prices exclude tools and long-context surcharges. OpenAI GPT-5.4 API model documentation Checked 2026-10-10 · high confidence	75.0%	Closed weightsLargePrice proxy	2026-03-05	Source ↗
12	Claude Opus 4.6Anthropic · Model Evaluation & model details Reported on Anthropic's Glasswing page. Maps the Glasswing OSWorld comparator to Claude Opus 4.6. Later revised-harness scores are not substituted. Harness code: uncertain. Price proxy: the size class comes from the standard token price of Claude Opus 4.6; parameters are undisclosed. Reported OSWorld evaluation Claude Opus 4.6 Closed weights · Large · Price proxy Token price: $5 input / $25 output per million. Blended 3:1: $10/M. Undisclosed parameters; size uses standard uncached token pricing. Tool charges excluded. Anthropic model pricing Checked 2026-10-10 · high confidence	72.7%	Closed weightsLargePrice proxy	2026-02-05	Source ↗
13	Claude Opus 4.5Anthropic · Model Evaluation & model details OSWorld-Verified self-reported result; reported on anthropic.com. Direct named-model computer-use evaluation. UiPath’s separate scaffold result is mapped separately rather than conflated with this baseline. Harness code: uncertain. Price proxy: the size class comes from the standard token price of Claude Opus 4.5; parameters are undisclosed. Reported OSWorld evaluation Claude Opus 4.5 Closed weights · Large · Price proxy Token price: $5 input / $25 output per million. Blended 3:1: $10/M. Closed hosted model with undisclosed parameter count. Published standard uncached first-party text prices are $5 input/$25 output per million tokens, giving a $10/M blend and Large price-proxy category. Cache, batch, and tool charges are excluded. Claude Opus 4.5 launch announcement Anthropic standard API pricing Checked 2026-10-10 · high confidence	66.3%	Closed weightsLargePrice proxy	2025-11-24	Source ↗
14	Kimi K2.5Moonshot AI · Model Evaluation & model details Self-reported in technical paper; GUI-only actions without external tools on OSWorld-Verified. Technical paper section 5.1.2 explicitly reports GUI-only OSWorld actions without external tools; no external runtime language model is included. Reuses the existing Kimi K2.5 checkpoint record. Harness code: not separately classified. Published size: from the parameter count of Kimi K2.5. Reported OSWorld evaluation Kimi K2.5 Open weights · Large · Published size · 1,000B total · 32B active Modified-MIT weights verified in public repository (64 safetensors shards). Swarm orchestration is distinct from these model weights. Price table was absent from retrieved pricing text and direct retrieval was blocked; token rates omitted. Moonshot Kimi K2.5 model card Kimi API pricing Checked 2026-10-10 · high confidence	63.3%	Open weightsLarge	2026-01-27	Source ↗
15	GLM-5V-TurboZhipu AI · Model Evaluation & model details Self-reported VLM result; reported on docs.z.ai. Named multimodal API model; exact parameter count and weight release remain unverified. Uses current standard provider token prices for the exact named model, excluding tool and cache fees. Harness code: uncertain. Price proxy: the size class comes from the standard token price of GLM-5V-Turbo; parameters are undisclosed. Reported OSWorld evaluation GLM-5V-Turbo Uncertain weights · Large · Price proxy Token price: $1.2 input / $4 output per million. Blended 3:1: $1.9/M. Exact multimodal API model is documented, but its parameter count and evaluated weights are not verified. The Z.AI pricing page currently omits this model; the serving Vercel AI Gateway directly lists its Z.AI route at standard $1.20 input/$4 output per million tokens (200K context), a $1.90/M blend. This is a provider price proxy, not a claim about parameter size. The exact zai-org checkpoint lookup returns HTTP 401 and the author search has no 5V release. GLM-5V-Turbo official model overview Vercel AI Gateway GLM-5V-Turbo Z.AI route pricing Z.AI current pricing; exact model absent Z.AI author checkpoint search Checked 2026-10-10 · medium confidence	62.3%	Uncertain weightsLargePrice proxy	2026-04-01	Source ↗
16	Claude Sonnet 4.5Anthropic · Model Evaluation & model details OSWorld-Verified, official framework, 100 max steps, 4-run avg; reported on anthropic.com. Named-model computer-use evaluation. The official launch and standard API rates establish identity and price category. Harness code: uncertain. Price proxy: the size class comes from the standard token price of Claude Sonnet 4.5; parameters are undisclosed. Reported OSWorld evaluation Claude Sonnet 4.5 Closed weights · Large · Price proxy Token price: $3 input / $15 output per million. Blended 3:1: $6/M. Closed hosted model with undisclosed parameter count. Published standard uncached first-party text prices are $3 input/$15 output per million tokens, giving a $6/M blend and Large price-proxy category. Cache, batch, and tool charges are excluded. This uses the standard context tier (up to 200K input tokens), not long-context premium pricing. Claude Sonnet 4.5 launch announcement Anthropic standard API pricing Checked 2026-10-10 · high confidence	61.4%	Closed weightsLargePrice proxy	2025-09-29	Source ↗
17	UiPath Screen Agent w/ GPT-5UiPath · Harness Evaluation & model details Official OSWorld-Verified workbook reports 53.63% with GPT-5 at 50 steps, rounded here to 53.6%; this is the September 2025 run. The official OSWorld-Verified workbook identifies the 53.63% (stored rounded 53.6%) result as UiPath Screen Agent w/ GPT-5, 50 steps, dated September 14, 2025. Opus 4.5 results in the same workbook are separate later runs. Exact historical scaffold release and auxiliary-model configuration are unverified; current UiPath evaluation code does not prove the older run was released. Harness code: uncertain. Price proxy: the size class comes from the standard token price of GPT-5; parameters are undisclosed. Official OSWorld-Verified results workbook Official benchmark and workbook host GPT-5 Closed weights · Large · Price proxy Token price: $1.25 input / $10 output per million. Blended 3:1: $3.4375/M. Parameter count is undisclosed. Large is a price proxy from standard uncached API text rates; it does not assert a physical parameter size. Tool fees and cached, batch, flex, and priority rates are excluded. OpenAI GPT-5 API model documentation Checked 2026-10-10 · high confidence	53.6%	Closed weightsLargePrice proxy	2025-09-14	Source ↗
18	Claude Haiku 4.5Anthropic · Model Evaluation & model details Self-reported result; reported on anthropic.com. The official launch identifies Haiku 4.5 and its computer-use result. Its $1/$5 standard rates put it in Large by the prescribed price proxy despite the Haiku product name. Harness code: uncertain. Price proxy: the size class comes from the standard token price of Claude Haiku 4.5; parameters are undisclosed. Reported OSWorld evaluation Claude Haiku 4.5 Closed weights · Large · Price proxy Token price: $1 input / $5 output per million. Blended 3:1: $2/M. Closed hosted model with undisclosed parameter count. Published standard uncached first-party text prices are $1 input/$5 output per million tokens, giving a $2/M blend and Large price-proxy category. Cache, batch, and tool charges are excluded. Claude Haiku 4.5 launch announcement Anthropic standard API pricing Checked 2026-10-10 · high confidence	50.7%	Closed weightsLargePrice proxy	2025-10-15	Source ↗
19	Agent S2 + Claude 3.7Simular AI · Harness Evaluation & model details Open-source modular agent; evaluated on 50-step OSWorld tasks. The paper pins Claude 3.7 Sonnet as manager/worker and UI-TARS-72B-DPO as visual grounding. Release v0.2.0 also uses text-embedding-3-small for uncached memory retrieval; include it conservatively. Tesseract OCR and UNO are conventional tools. Optional Perplexica is disabled in the release CLI and is not asserted as a required evaluated model. The open harness thus has mixed model weights and Large largest-component size. Harness code: open. Price proxy: the size class comes from the standard token price of Claude 3.7 Sonnet; parameters are undisclosed. Size reflects the largest of 3 models; parameters are not summed. Reported OSWorld evaluation Agent S2 original launch result Agent S2 paper, section 4.1 Agent S2 release v0.2.0 Agent S2 v0.2.0 CLI optional search configuration Claude 3.7 Sonnet Closed weights · Large · Price proxy Token price: $3 input / $15 output per million. Blended 3:1: $6/M. Closed hosted model with undisclosed parameter count. Published standard uncached first-party text prices are $3 input/$15 output per million tokens, giving a $6/M blend and Large price-proxy category. Cache, batch, and tool charges are excluded. The launch announcement supplies historical rates; current pricing tables omit this retired model. Thinking tokens are billed as output. Claude 3.7 Sonnet launch announcement Checked 2026-10-10 · high confidence UI-TARS-72B-DPO Open weights · Medium · Published size · 73.406B total Public Apache-2.0 visual grounding checkpoint used by Agent S2. The model is published as 72B; its inventory includes about 73.4B tensors, including the vision stack, across 31 safetensor shards. Both totals are Medium. Keep this original DPO release distinct from later UI-TARS-1.5. ByteDance-Seed/UI-TARS-72B-DPO model card and released files ByteDance-Seed/UI-TARS-72B-DPO checkpoint inventory Agent S2 paper, section 4.1 Checked 2026-10-10 · high confidence OpenAI text-embedding-3-small Closed weights · Small · Estimated Agent S2 v0.2.0 hard-codes this model for narrative/episodic memory retrieval when embeddings are not cached. Official embedding price is $0.02 per million input tokens. It has no generative output-token rate, so the 3:1 text-generation price formula does not apply and output price is not fabricated as zero. Small is a provisional low-cost embedding-service category; parameter count is undisclosed. text-embedding-3-small official model documentation Agent S2 v0.2.0 embedding engine Agent S2 v0.2.0 retrieval implementation Checked 2026-10-10 · medium confidence	34.5%	Mixed weightsLargePrice proxyLargest component	2025-03-12	Source ↗
20	OpenAI Operator (CUA)OpenAI · System Evaluation & model details Original January 2025 CUA; Agent S2 reports the 32.6% baseline at 50 steps. The OpenAI launch score used a different step budget. Original January 2025 CUA evaluated at 50 steps; Agent S2’s published baseline table corroborates the 32.6 setting. The broader launch page also reports 38.1 with a different step budget. Do not map this result to later Operator updates. Product-side safety monitors are not claimed as components of the benchmark policy absent evidence. Harness code: closed. Estimated for OpenAI original Computer-Using Agent (January 2025): no verified parameter count or standard token price, so the size class is provisional. Reported OSWorld evaluation Agent S2 baseline table with 50-step original CUA OpenAI original Computer-Using Agent (January 2025) Closed weights · Large · Estimated The original Operator CUA combines GPT-4o vision with reinforcement-learned reasoning. Its parameter count and January evaluation token prices are undisclosed. Large is a provisional category informed by the GPT-4o-based frontier architecture and later computer-use-preview pricing, not a parameter estimate or assertion that the March API snapshot is identical. Do not substitute later o3-based Operator or computer-use-preview-2025-03-11 for this historical model. OpenAI original CUA announcement and evaluation Later computer-use-preview model and dated snapshot; comparative evidence only Checked 2026-10-10 · medium confidence	32.6%	Closed weightsLargeEstimated	2025-01-23	Source ↗
21	Qwen3-VL-235B-A22B-InstructAlibaba · Model Evaluation & model details Self-reported OSWorld result from Qwen3-VL technical report Table 2 (Instruct variant). The separate OSWorldG grounding score is 66.7%; it is not an OSWorld task-success result. Exact Instruct checkpoint identified by Table 2 of the primary technical report. Corrected a previously imported OSWorldG grounding score to the matching 31.6 OSWorld result; Thinking is a distinct evaluated variant and is not substituted. Official downloadable Instruct weights verified; exact evaluation harness release remains unverified. Harness code: uncertain. Published size: from the parameter count of Qwen3-VL-235B-A22B-Instruct. Qwen3-VL technical report, Table 2 (page 15) Qwen3-VL-235B-A22B-Instruct model card and released files Qwen3-VL-235B-A22B-Instruct Open weights · Large · Published size · 235B total · 22B active The official report specifies 235B total and 22B active parameters. Its Table 2 distinguishes the Instruct model's 31.6 OSWorld score from its 66.7 OSWorldG grounding score. The exact Instruct checkpoint inventory is public and ungated, with 96 safetensors shards at revision 710c13861be6c466e66de3f484069440b8f31389. Stored tensors total 235.670022896B including auxiliary components; the published architectural total determines the unchanged Large category. Qwen3-VL technical report, Table 2 (page 15) Qwen3-VL-235B-A22B-Instruct model card and released files Qwen3-VL-235B-A22B-Instruct checkpoint inventory Checked 2026-10-11 · high confidence	31.6%	Open weightsLarge	2025-11-26	Source ↗
22	Qwen2.5 VL 72B InstructAlibaba Cloud / Qwen Team · Model Evaluation & model details Self-reported result; reported on huggingface.co. Exact named instruction-tuned checkpoint; model card and public safetensor inventory verified. No separate learned runtime helper is identified. Harness code: not separately classified. Published size: from the parameter count of Qwen2.5-VL-72B-Instruct. Reported OSWorld evaluation Qwen2.5-VL-72B-Instruct Open weights · Medium · Published size · 73.411B total The official card identifies the instruction-tuned 72B model and includes the OSWorld result. Public checkpoint inventory verifies 38 safetensor shards and approximately 73.4108B stored parameters including vision components. This full inventory drives the total; it falls in the same category as the nominal language-model size. The 72B release uses the Qwen license; 32B uses Apache-2.0. Open weights is not an assertion of identical license terms. Qwen/Qwen2.5-VL-72B-Instruct model card and released files Qwen/Qwen2.5-VL-72B-Instruct checkpoint inventory Checked 2026-10-10 · high confidence	8.8%	Open weightsMedium	2025-01-26	Source ↗
23	Qwen2.5 VL 32B InstructAlibaba Cloud / Qwen Team · Model Evaluation & model details Self-reported result; reported on huggingface.co. Exact named instruction-tuned checkpoint; model card and public safetensor inventory verified. No separate learned runtime helper is identified. Harness code: not separately classified. Published size: from the parameter count of Qwen2.5-VL-32B-Instruct. Reported OSWorld evaluation Qwen2.5-VL-32B-Instruct Open weights · Small · Published size · 33.453B total The official card identifies the instruction-tuned 32B model and includes the OSWorld result. Public checkpoint inventory verifies 18 safetensor shards and approximately 33.4527B stored parameters including vision components. This full inventory drives the total; it falls in the same category as the nominal language-model size. The 72B release uses the Qwen license; 32B uses Apache-2.0. Open weights is not an assertion of identical license terms. Qwen/Qwen2.5-VL-32B-Instruct model card and released files Qwen/Qwen2.5-VL-32B-Instruct checkpoint inventory Checked 2026-10-10 · high confidence	5.9%	Open weightsSmall	2025-03-24	Source ↗

About this benchmark

OSWorld evaluates multimodal computer-use agents in real desktop environments across 369 tasks involving web apps, desktop software, files, and workflows spanning multiple applications.

It is valuable for teams building GUI agents because tasks require visual grounding, keyboard and mouse execution, OS knowledge, and error recovery, not only text planning.

Modern reports often distinguish original OSWorld, OSWorld-Verified, and submitter-run variants; read source details before comparing human-level claims.

Self-reported and independently verified rows coexist; setup differences can matter as much as the model.

Example tasks

Three public tasks quoted from benchmark sources:

"Can you enable the 'Do Not Track' feature in Chrome to enhance my online privacy?" Citation: OSWorld example JSON
"Can you make my computer bring back the last tab I shut down?" Citation: OSWorld example JSON
"Computer, please navigate to the area in my browser settings where my passwords are stored. I want to check my login information for Etsy without revealing it just yet." Citation: OSWorld example JSON

Methodology

Original OSWorld uses execution-based validators that check final computer state after the agent acts in configured VM environments.
Reported metric is success rate; the original paper reported a 72.36% human baseline and 12.24% for the best early model.
OSWorld-Verified adds independent or standardized re-runs for some systems; self-reported rows can use different max steps, OS images, and tool permissions.
We track public results with source URLs and note whether the source claims verified or independent execution.

Related benchmarks

Compare this benchmark with related pages from the hub:

osworld-2 webarena webvoyager online-mind2web

Back to benchmark hub

Frequently asked questions

Which system is currently best on OSWorld? +

Qwen3.8-Max is the system/agent setup currently leading with a tracked score of 86.1%. This ranking reflects submitted system setups (model plus tools and policy), not just a base model. Based on our latest tracked results, last updated Oct 11, 2026.

What should I read into a OSWorld score? +

OSWorld scores are most useful for within-benchmark ranking. Read the Notes column to understand setup context, and use the methodology section before making procurement or architecture decisions.

Are these independently verified? +

Not always. Some rows are independently benchmarked and some are team-reported. Use each source link and notes field to verify evidence level before drawing strong conclusions.

Can I compare every row directly? +

Use caution. Rows can vary by evaluator, harness, attempt budget, tool access, task filtering, or verification level. Source links and notes are part of the score, not decoration.

OSWorld scores

Results

Qwen3.8-Max (August 2026 launch)

Claude Mythos Preview

Claude Mythos 5

Claude Fable 5

Claude Opus 4.8

MiMo-V2.6-Pro-RL

Claude Sonnet 5

MiMo-V2.6-Flash-RL

Claude Sonnet 4.6

OSAgent (October 2025 undisclosed policy)

GPT-5.4

Claude Opus 4.6

Claude Opus 4.5

Kimi K2.5

GLM-5V-Turbo

Claude Sonnet 4.5

GPT-5

Claude Haiku 4.5

Claude 3.7 Sonnet

UI-TARS-72B-DPO

OpenAI text-embedding-3-small

OpenAI original Computer-Using Agent (January 2025)

Qwen3-VL-235B-A22B-Instruct

Qwen2.5-VL-72B-Instruct

Qwen2.5-VL-32B-Instruct

About this benchmark

Example tasks

Methodology

Links

Related benchmarks

Frequently asked questions