By Aijaz Alam
How often do AI agents succeed at real tasks?
Top AI agents succeeded on 77.3% of real-world tasks in Terminal-Bench in early 2026, up from 20% in 2025, and on about 66% of real computer tasks in OSWorld, according to the Stanford AI Index 2026. That still means about one failed attempt in three on OSWorld, where people in the original 2024 study completed over 72.36% of tasks.
- 77.3%Agent success on real-world tasks in Terminal-Bench, up from 20% in 2025Stanford AI Index 2026
- About 66%Agent success on real computer tasks in OSWorld, up from 12%Stanford AI Index 2026
- 72.36%OSWorld tasks people completed in the benchmark's 2024 studyOSWorld paper, arXiv, April 2024
- 93%Cybersecurity problems solved by AI agents, up from 15% in 2024Stanford AI Index 2026
- About 17 hoursLength of software task the most capable model tested finishes half the timeMETR, model released Apr 2026; METR calls measurements above 16 hours unreliable
The numbers
The full breakdown
AI agent success on benchmark tasks (Stanford AI Index 2026)
| Benchmark | What it tests | Earlier | Latest (2026 AI Index) |
|---|---|---|---|
| Terminal-Bench | Real-world tasks in a computer terminal | 20% (2025) | 77.3% |
| OSWorld | Real computer tasks across operating systems | 12% | About 66% |
| Cybersecurity agent tasks | Solving cybersecurity problems | 15% (2024) | 93% |
| SWE-bench Verified | Fixing real software issues (coding) | 60% | Near 100% a year later |
| Real household tasks (robots) | Folding clothes, washing dishes and similar | Not given | 12% |
Best scores reported in the AI Index 2026 (13 Apr 2026). Benchmark scores are lab tests of the best systems, not the success rate of agents deployed in businesses. The AI Index does not give the start year for the OSWorld 12%.
OSWorld when it was launched: people against the best model (April 2024)
| Who | Share of OSWorld tasks completed |
|---|---|
| People | Over 72.36% |
| Best AI model tested | 12.24% |
OSWorld paper (arXiv 2404.07972, first submitted 11 Apr 2024). The models mainly struggled with recognising on-screen elements (GUI grounding) and operational knowledge. Two years later the AI Index puts the best agents at about 66%, close to the human level.
Over time
How long a software task the most capable AI agents can finish half the time (METR time horizon)
| Model release | Model | 50% time horizon (human expert time) | 80% time horizon |
|---|---|---|---|
| Mar 2023 | GPT-4 | About 4 minutes | About 1 minute |
| May 2024 | GPT-4o | About 7 minutes | About 1 minute |
| Dec 2024 | o1 | About 39 minutes | About 7 minutes |
| Feb 2025 | Claude 3.7 Sonnet | About 1 hour | About 12 minutes |
| Apr 2025 | o3 | About 2 hours | About 30 minutes |
| Aug 2025 | GPT-5 | About 3 hours 23 minutes | About 38 minutes |
| Nov 2025 | Claude Opus 4.5 | About 4 hours 53 minutes | About 49 minutes |
| Dec 2025 | GPT-5.2 | About 5 hours 52 minutes | About 66 minutes |
| Feb 2026 | Claude Opus 4.6 | About 12 hours | About 70 minutes |
| Apr 2026 | Claude Mythos Preview (early) | About 17 hours (METR: measurements above 16 hours are unreliable) | About 3 hours 6 minutes |
METR Time Horizon 1.1 estimates (metr.org/time-horizons), a selection of the models METR marks as the state of the art when released, with model names as METR writes them; rounded from METR's point estimates in minutes. The 50% horizon is the length of task, measured by how long it takes a human expert, at which the agent is predicted to succeed half the time. Confidence intervals are wide (Claude Opus 4.6: about 5 to 61 hours), and METR says measurements above 16 hours are unreliable with its current task suite, which applies to the Claude Mythos Preview (early) figure. Software tasks only. METR stopped actively updating the page on 8 Sep 2026.
Why sources differ
Same question, different answers
| Number | When | Who measured | Why it differs |
|---|---|---|---|
| 77.3% | 2026 | Stanford AI Index (Terminal-Bench) | Share of a fixed set of terminal tasks the best agent completes. |
| About 66% | 2026 | Stanford AI Index (OSWorld) | Real computer tasks with a graphical interface; still below the 72.36% people reached. |
| 93% | 2026 | Stanford AI Index (cybersecurity) | A narrower set of cybersecurity problems. |
| 12% | 2026 | Stanford AI Index (household robots) | Physical tasks done by robots, which remain far harder. |
| 50% at about 17 hours | Apr 2026 | METR time horizon | Not a success rate: the length of task an agent finishes half the time, on software tasks. |
There is no single AI agent success rate, because each benchmark is a different set of tasks. Terminal-Bench gives an agent jobs in a command-line terminal, OSWorld gives it a full desktop with windows and menus, and the cybersecurity tests use a narrower kind of problem. Scores also depend on which model and which agent software is tested; the AI Index reports the best results, not typical ones.
METR measures something else: instead of one percentage, it asks how long a task, timed by a human expert, an agent can complete with 50% or 80% reliability. In its original March 2025 study, models succeeded almost every time on tasks that take people under 4 minutes but less than 10% of the time on tasks over about 4 hours. Success falls with task length, so a high score on short tasks says little about long projects.
None of these is the success rate of agents deployed at work, where tasks are less tidy and need context the agent does not have. No source in this data measures that directly.
What it means
How to use these numbers
Agents have improved faster than almost any other AI capability. On Terminal-Bench the best score went from 20% to 77.3% in about a year, and on OSWorld from 12% to about 66%, close to the 72.36% people managed when the test was built. METR's measure tells the same story in time: the task length the best models finish half the time went from about 4 minutes in 2023 to several hours in 2025.
The catch is reliability. Even the best OSWorld agent fails about one attempt in three, and every model's 80% time horizon is far shorter than its 50% one: tasks it can finish half the time it cannot finish reliably. For a business, that means agents fit work where a failure is cheap to spot and redo, with a person checking the result, rather than unattended jobs where a silent mistake is costly.
For journalists: name the benchmark and the date with any agent success rate, and do not present a benchmark score as the share of real-world tasks agents complete.
Method and sources
How it was measured
Method
- Publishers
- Stanford Institute for Human-Centered AI, AI Index 2026 (13 Apr 2026); OSWorld authors (arXiv 2404.07972); METR, Time Horizon 1.1
- What is measured
- Benchmarks: share of a fixed task set completed by the best tested agent. METR: task length (human expert time) at which an agent is predicted to succeed 50% or 80% of the time
- Human baseline
- OSWorld: people completed over 72.36% of tasks in the 2024 study
- Limits
- Lab tests of frontier models with set agent software; not adoption and not success in deployed business use. METR covers software tasks only and calls measurements above 16 hours unreliable
- Checked at source
- 31 Oct 2026 on hai.stanford.edu, arxiv.org and metr.org
- Next release
- The AI Index is published each spring; METR's time-horizon page has not been actively updated since 8 Sep 2026
- Last changed
- 31 Oct 2026
- Data files
- CSV · JSON (stable links)
Sources
- AI Index Report, Stanford HAI Research
- arXiv preprint server, Cornell University (arXiv) Study
- METR research notes, METR (Model Evaluation and Threat Research) Study
Cite this page
APA
MLA
HTML link
FAQ
More questions people ask
What is an AI agent's time horizon?
METR's measure of the length of task, timed by how long it takes a human expert, that an agent is predicted to complete with a given reliability. A 2-hour 50% horizon means a 2-hour task is finished about half the time.
Are AI agents better than people at computer tasks?
Not yet on OSWorld: people completed over 72.36% of its tasks in the 2024 study, and the best agents reached about 66% in the 2026 AI Index. On narrower tests such as Terminal-Bench (77.3%) agents score higher, but there is no human figure from the same source.
Why do agents fail on long tasks?
Success drops as tasks get longer. In METR's March 2025 study, models succeeded almost always on tasks taking people under 4 minutes but less than 10% of the time on tasks over about 4 hours.
Do these scores show how agents perform at work?
No. They are lab benchmarks run on the best systems. The AI Index itself calls the OSWorld result a structured benchmark, and real work needs context and tools a benchmark does not give.
Every number on this page was checked at its original source (Stanford HAI, Cornell University (arXiv), METR (Model Evaluation and Threat Research)). Free to cite with a link to this page.