ResearchAll sizesGlobal
77.3%
success rate for AI agents on real-world tasks in Terminal-Bench, up from 20% in 2025 (Stanford AI Index 2026)
ResearchUpdated 23 Oct 2026
About 66%task success for AI agents on OSWorld, a test of real computer tasks, up from 12%; agents still fail about 1 in 3 attempts
About 66% task success for AI agents on OSWorld, a test of real computer tasks, up from 12%; agents still fail about 1 in 3 attempts (2025-2026, Stanford HAI, AI Index Report)
ResearchChecked at source, 23 Oct 2026All sizesGlobalAI agents · Capability
AI agent success on benchmark tasks (Stanford AI Index 2026)
| Group | Earlier | Latest |
|---|---|---|
| OSWorld (computer tasks) | 12% | 66% |
| Terminal-Bench (real-world tasks) | 20% | 77.3% |
In plain words
What this number really says
About 66% task success for AI agents on OSWorld, a test of real computer tasks, up from 12%; agents still fail about 1 in 3 attempts (2025-2026).
About 66%
the latest figure (2025-2026)
Research
type of source
Why you can trust it. Stanford HAI, AI Index Report: research, research firm or analyst compilation.
Key facts
- Period 2025-2026
- Region Global
- Company size All sizes
- Next data Yearly
Source: Stanford HAI, AI Index Report, 2025-2026
Method and use
How it was measured, and how to use it
How it was measured
- Publisher
- Stanford HAI, AI Index Report
- Source type
- Research: Research firm or analyst compilation
- What it measures
- About 66% task success for AI agents on OSWorld, a test of real computer tasks, up from 12%; agents still fail about 1 in 3 attempts
- Data period
- 2025-2026 (ends 31 Mar 2026)
- Published by the source
- 13 Apr 2026
- In the source's words
AI agents made a leap from 12% to ~66% task success on OSWorld, which tests agents on real computer tasks across operating systems
- Checked at source
- 23 Oct 2026: Read on hai.stanford.edu (2026 AI Index Report, top takeaways) on 2026-10-23. Benchmark score, not real-world adoption.
- Reuse
- CC BY-ND 4.0 for the report; check before reuse. Compiles data from many sources; cite the AI Index and its original source.
- Next expected update
- Yearly (Stanford HAI, yearly (April))
- Original page
- hai.stanford.edu: AI Index Report
- Data files
- CSV · JSON (stable links)
Full change log
No changes recorded yet. First checked at the source on 23 Oct 2026.
Cite this number
APA
Stanford HAI. (2026). About 66% task success for AI agents on OSWorld, a test of real computer tasks, up from 12%; agents still fail about 1 in 3 attempts. Via digitaltreed, https://digitaltreed.com/statistics/ai-agents/agent-success-osworld
MLA
"About 66% task success for AI agents on OSWorld, a test of real computer tasks, up from 12%; agents still fail about 1 in 3 attempts." Stanford HAI, AI Index Report, 2025-2026. digitaltreed, https://digitaltreed.com/statistics/ai-agents/agent-success-osworld.
HTML link
<a href="https://digitaltreed.com/statistics/ai-agents/agent-success-osworld">About 66% task success for AI agents on OSWorld, a test of real computer tasks, up from 12%; agents still fail about 1 in 3 attempts</a> (Stanford HAI, 2025-2026)
Embed the chart
Keep the source link under the chart. The chart updates when the number does.
FAQ
Questions about this number
What is the latest figure for capability (Global)?
About 66% task success for AI agents on OSWorld, a test of real computer tasks, up from 12%; agents still fail about 1 in 3 attempts, for 2025-2026, according to Stanford HAI, AI Index Report. We checked it at the source on 23 Oct 2026.
Put this number to work
We help small and mid-size firms turn data like this into useful content, answers for customers and visibility in AI search.
Every number on this page was checked at its original source. Free to cite with a link to this page.