LabPolicy

Evaluations

The public evidence on what frontier agents can do and where they fail, drawn as leaderboards, with the people who measured it credited on every card.

The evidence we track

Five measurements, each from its publisher's latest release. We build our environments against them.

Agents · 15 models

Time horizons, every model

METR measures how long a task an AI agent can complete, counted in the time the same task takes a skilled human. A model's time horizon is the length of task it finishes half the time.

It is the clearest single measure of what agents can be trusted to do alone, and so of how much has to be learned in training before they are: the longer an agent works without a person checking, the more its environment and its reward decide what it does.

Each point is the state-of-the-art model at its release; confidence intervals are wide. Bars on a log scale. Data from METR, Task-Completion Time Horizons of Frontier AI Models, 8 May 2026.

  1. Claude Mythos PreviewNewest17.4 hours
  2. Claude Opus 4.612.0 hours
  3. GPT-5.25.9 hours
  4. Claude Opus 4.54.9 hours
  5. GPT-53.4 hours
The 10 models before them
  1. o32.0 hours
  2. Claude 3.7 Sonnet1.0 hours
  3. o139 min
  4. o1-preview20 min
  5. Claude 3.5 Sonnet11 min
  6. GPT-4o7 min
  7. GPT-44 min
  8. GPT-3.5 Turbo Instruct36 sec
  9. davinci-0028 sec
  10. GPT-23 sec

Safeguards · 8 risks

The mitigation gap

Share of organisations that actively mitigate each AI risk, beside the share that call it relevant, widest gaps first. 43.8% call autonomous system actions a relevant risk; 28.8% mitigate it.

Data from Stanford HAI, AI Index 2026.

  1. Reputation22.0%of 37.3%
  2. Autonomous system actions28.8%of 43.8%
  3. IP infringement39.7%of 51.4%
  4. Personal privacy44.0%of 54.4%
  5. Cybersecurity61.5%of 71.8%

Incidents · 12 years

Reported AI incidents

Incidents reported to the AI Incident Database each year. The count has more than quadrupled since 2021.

Data from AI Incident Database, via Stanford HAI.

  1. 2025362
  2. 2024281
  3. 2023168
  4. 2022104
  5. 202177

How we evaluate an agent

The same environments an agent trains in are the ones it is tested in, before deployment and after.

  1. 01Specify

    Every environment starts from the rules an agent will be held to: the statute, the regulation, the permissions of the system it runs in. Members in law write them as tests a program can check.

  2. 02Build

    Members in computation build the environment: the tools, files, accounts and counterparties an agent meets on the job, in a sandbox where every action is recorded and nothing leaves.

  3. 03Reward

    The reward is where a guardrail becomes something the agent learns. Ours pay for finishing the task within the rules, and pay nothing for a result reached by breaking one.

  4. 04Evaluate

    We evaluate agents in the same environments before and after deployment, publish the results with the evidence behind them, and brief the legislators who ask what a rule would cost.