Benchmark Almanac

OSWorld-G

Fara1.5-27B 71.3% leads since Jun 2026 Fading

Models map 564 desktop instructions to the right screen element or point, covering text matching, element recognition, layout understanding and fine manipulation.

Released with the Jedi grounding dataset. NeurIPS 2025 spotlight. Scored by click accuracy.

Made by XLANG Lab , University of Hong Kong (OSWorld team) Released May 2025 Finding things on screen All tasks public Last in a launch post Jan 2026 , dropped before it saturated Last model added Jun 2026
Level
Everyday Most adults, without training
Scored
Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
Answers
Short answer A number, word or short string matched to a key.
xlang-ai/OSWorld-G 184 stars Paper Dataset
5
Models scored in lab cards
4
Launch posts list it, from 4 labs
5
System cards report it, from 5 labs

Fara1.5-27B leads with 71.3% among models released by May 2026

Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.

The newest model with a score here came out in May 2026, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 71.3 Fara1.5-27B
  2. 68.3 Qwen3-VL-235B-A22B
  3. 59.0 STEP3-VL-10B
  4. 53.2 InternVL3.5-241B-A28B
  5. 52.5 Kimi-VL-A3B-Thinking-2506
  1. 1 Fara1.5-27B 71.3%
  2. 2 Qwen3-VL-235B-A22B 68.3%
  3. 3 STEP3-VL-10B 59%
  4. 4 InternVL3.5-241B-A28B 53.2%
  5. 5 Kimi-VL-A3B-Thinking-2506 52.5%
#ModelScoreSettingDateSource
1 Fara1.5-27B 71.3% Jun 2026 Model card
2 Qwen3-VL-235B-A22B 68.3% Sep 2025 Model card
3 STEP3-VL-10B 59% Jan 2026 Model card
4 InternVL3.5-241B-A28B 53.2% Aug 2025 Model card
5 Kimi-VL-A3B-Thinking-2506 52.5% Jun 2025 Model card

4 models have held number one since its release in May 2025

Lab claims count from the day the card came out, independent runs from the model’s release.

Moonshot AI Shanghai AI Lab Qwen Microsoft
Lab claims Not tested since Jun 2026
Fara1.5-27B 71.3%

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Fara1.5-27B Qwen3-VL-235B-A22B InternVL3.5-241B-A28B 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Fara1.5-27B InternVL3.5-241B-A28B

Labs headline it less often now, last in Jan 2026

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

Moonshot AI
Shanghai AI Lab
Qwen
StepFun
DateLaunch postScoreAs the post writes it
13 Jan 2026 STEP3-VL-10B 59 OSWorld-G
23 Sep 2025 Qwen3-VL-235B-A22B 68.3 OSWorldG
25 Aug 2025 InternVL3.5-241B-A28B 53.2 OSWorld-G
21 Jun 2025 Kimi-VL-A3B-Thinking-2506 52.5 OSWorld-G (Acc, full set with refusal)

5 system cards report it, from 5 labs

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
18 Jun 2026 Fara1.5-27B technical report 71.3 OSWorld-G Refined grounding accuracy (Table 6) Capabilities
14 Jan 2026 STEP3-VL-10B technical report 59 OSWorld-G, SeRe (Table 1) Capabilities
26 Nov 2025 Qwen3-VL-235B-A22B technical report 68.3 OSWorldG (Table 2) Capabilities
25 Aug 2025 InternVL3.5-241B-A28B technical report 53.2 OSWorld-G (Table 2, Table 10) Capabilities
23 Jun 2025 Kimi-VL-A3B-Thinking-2506 technical report 52.5 OSWorld-G (Acc), full set with refusal samples (Table 5, section 4.3) Capabilities

More in Computer and browser use

All 56 in Computer and browser use
  • BrowseComp In use Hard-to-find facts that need persistent web browsing. Short verifiable answers graded for exact correctness. Claude Fable 5 93.3%
  • ScreenSpot-Pro In use GUI grounding on high-resolution professional-software screenshots. The model must locate the target UI element. GPT-6 Astra 92.7%
  • AndroidWorld In use Agents complete 116 tasks across 20 real Android apps in an emulator. Each task checks the device's system state to judge success. GPT-6 Astra 100%
  • OSWorld 2.1 In use Agents complete 108 long-horizon workflows across 31 self-hosted websites and desktop apps, averaging 318 tool calls. Scored by binary completion and checkpoint partial credit. Claude Opus 5.5 81.8%
  • WebArena-Verified In use Agents complete 812 audited WebArena tasks on self-hosted websites. Deterministic evaluators check the agent's answer and captured network traces. Muse Spark 1.1 69%
  • MobileWorld In use Agents complete 201 tasks across 20 mobile apps, including tasks that need user clarification or MCP tools. Backend checks verify success. Kimi K3 74.4%