The Threshold Report

AI progress, only when it's real.

Mode:·

Issue #3 · June 2, 2026

Text 117 (+10) · Audio 109 (+2) · Video 120 (+5) · Image 100 (+0) · Agents 112 (+5) · Robots 100 (+0) · Science 100 (+0) · Other 100 (+0)

What became real

Text+10Jun 2

Sample: day-long coding tasks completed end to end in a public coding agent

A publicly available coding agent now completes multi-hour software tasks from a spec. Small teams can hand off whole features, not just snippets.

What still breaks: Ambiguous specs still produce confidently wrong code.
SOURCES · example.com
Agents+5Jun 2

Sample: vision-grounded agents read screenshots and act on them in a shipped product

A public agent now reliably reads on-screen content and takes the corresponding action. Software help becomes 'do it for me', not 'tell me how'.

What still breaks: Dense dashboards and tiny controls still confuse it.
SOURCES · example.com
Video+5Jun 2

Sample: physics realism improves measurably on a tracked video benchmark

A shipped video model posted a clear gain on a public benchmark for motion and physics. Fewer uncanny artifacts in everyday generated clips.

What still breaks: Object permanence across cuts is still weak.
SOURCES · example.com
Audio+2Jun 2

Sample: studio-quality dubbing expands to ten more languages

An existing dubbing product added broad language coverage. Creators can localize content for new markets cheaply.

What still breaks: Regional accents within a language remain hit-or-miss.
SOURCES · example.com

Not real yet

Announced or reported, but nothing you can use or verify. Some of these will become real. Some won’t.

AudioJun 2

Sample: reports of a real-time singing-voice translator (unconfirmed)

Waiting on: Primary verification

example.com
VideoJun 2

Sample: research demo shows ten-minute coherent video (not yet released)

Waiting on: Not yet publicly usable

example.com