Skip to content
Humanopia
Safety

Labs agree to share evaluation results before launch

A voluntary agreement covers dangerous capability tests. Critics want it to be binding.

Humanopia Editorial Team · Oct 3, 2026 · 1 min read

The timing is not an accident. Competitors have shipped similar updates in recent months, and every lab wants to be the default choice when a company signs a yearly contract.

The practical advice is the same as always: try it on a low-stakes task first, check the output against a source you trust, and keep a human in charge of anything that matters.

What critics say

If you use these tools at work, ask your administrator whether the new feature is turned on, and what the company policy says about using it with customer data.

  • Rolling out region by region
  • Independent tests are still limited
  • Older versions stay available for now
  • Available first to paid plans

How we reported this

Independent researchers were quick to test the claims. Early results broadly match what the company said, with the usual caveat: benchmarks measure narrow skills, and real work is messier than a test set.

Stay ahead without
living inside the feed.

Get our weekly newsletter with the most important AI news, practical guides and thought-provoking essays, curated for curious, forward-thinking people.

Weekly, not daily Curated and practical No spam, ever