cd /news/ai-research/metrs-time-horizon-plot-ai-coding-so… · home › topics › ai-research › article
[ARTICLE · art-173405] src=dev.to ↗ pub= topic=ai-research verified=true sentiment=· neutral

METR’s Time Horizon Plot: AI Coding Soars, Judgment Lags

METR published an updated time horizon plot showing that AI models' autonomous task completion on software engineering benchmarks has doubled roughly every seven months since 2021, with models now handling tasks lasting over an hour versus seconds in 2021. The nonprofit's update notes that this exponential growth applies mainly to tasks with clear right answers like coding, while judgment-based tasks show shorter and more slowly improving time horizons.

recommendation

Automate coding tasks with clear pass/fail checks now, but keep humans reviewing work that depends on subjective judgment, since METR's data shows those horizons lag.

more in playbook: Evals and testing for LLM apps
by read1 min views9 publishedNov 20, 2026

The most misunderstood graph in AI just got a major update—and it reveals a crucial gap: AI models are getting exponentially better at coding, but they still suck at tasks requiring human judgment.

METR, the AI evaluation nonprofit, published its latest time horizon plot, which measures how long an AI can autonomously complete tasks before failing. The plot shows that for software engineering benchmarks, the time horizon has been doubling roughly every seven months since 2021. That’s exponential growth. A model that could only handle tasks lasting a few seconds in 2021 can now handle tasks lasting over an hour.

Most people look at that curve and think, “AI is taking over.” But the update includes a critical nuance: the plot only tracks tasks that have a clear right answer. Coding tasks fit that mold—compile or fail. Tasks requiring judgment, like deciding whether a news article is biased, are harder to automate because the answer is subjective.

METR specifically calls this out: models aren’t quite as good at tasks that require judgment as they are at coding. The time horizon for judgment-based tasks is shorter and improving more slowly. This isn’t a minor detail—it’s the difference between an AI that can write your unit tests and one that can review your PR for design flaws.

The exponential curve is real, but it’s not uniform. Coding is the low-hanging fruit. Judgment is the hard wall. If you’re planning your AI stack, invest in coding automation now, but keep humans in the loop for any task that requires nuance. The graph update doesn’t change that reality—it just makes it measurable.

Source: MIT Technology Review

── more in #ai-research 4 stories · sorted by recency
── more on @metr 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/metrs-time-horizon-p…] indexed:0 read:1min 2026-11-20 · —