By 2030, AI systems will be able to automate many real-world software tasks that take humans a month.
Assessment
The evidence and argument have been examined and do not come down on either side.
This is a forecast. Its main support is METR's measurements of the "time horizon" of AI agents: the length of software task, measured in human working time, that an agent completes with 50% success. That horizon has grown roughly exponentially, doubling about every seven months from 2019 to 2025 and somewhat faster in 2024–2025. By late 2025 frontier models were at several hours. A month of work is about 167 hours. If the doubling pace continues, five or six more doublings would reach month-long tasks between roughly 2028 and 2030, near the edge of the claim's deadline.
The extrapolation is uncertain for three reasons. First, the benchmark tasks are cleaner than real-world software work, and it is unsettled whether benchmark horizon trends generalize to real-world work. There is also some evidence that progress is slower on messier tasks. Second, a 50% success rate falls well short of reliably "automating" a task. Third, the trend could slow because of limits on compute, data, or long-running agent coherence, or speed up if AI begins accelerating AI research. Taken together, these make the claim about as likely to fail as to hold. The question would be settled by horizon measurements through 2027–2028 on tasks closer to real-world conditions.
Full reasoning: the evidence and decisions behind this verdict
This assessment draws on prior knowledge of METR's "Measuring AI Ability to Complete Long Tasks" (Kwa et al., March 2025, arxiv.org/abs/2503.14499). That paper found a horizon doubling time of about 7 months over 2019–2025, possibly about 4 months in 2024–2025. It extrapolated month-long horizons for roughly 2028–2031, and it noted that its tasks are less messy than real work. Later measurements put frontier models at a few hours by late 2025.
The arithmetic: going from about 5 hours to about 167 hours is a factor of about 33, or about 5 doublings. At 4–7 months per doubling, that takes 20–35 months from late 2025, which lands around 2027.5–2029.
Several discounts pull the other way: - External validity: on messier tasks, success and horizons were lower. - Threshold: "automate many" tasks plausibly needs more than 50% reliability, and the horizon at 80% success is about 4–5 times shorter. - The trend may break.
The argument from trend extrapolation is rated likely if its premises hold, but both of its central premises (that the trend persists and that it transfers to real work) are unassessed and genuinely uncertain. The argument from messy real-world tasks is likewise rated likely, with premises that are partly supported but not established. Neither argument dominates, so credence sits just under even. This pass ran no new web searches. A pass that pulls the latest METR horizon measurements from 2026 would sharpen the estimate.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Because frontier 50% time horizons have grown roughly exponentially, reaching several hours by late 2025, and that doubling pace will continue, about five more doublings would bring horizons to roughly 167 working hours (one month) by 2028–2030. Given that benchmark horizon trends generalize to real-world software work, AI would then automate many real-world month-long software tasks by 2030.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 doubtful1 inconclusive1 not separately assessed
Granting its premises, the conclusion becomes much more likely but is not guaranteed. A 50% success horizon on month-long tasks is weaker than routinely "automating" them, and the arithmetic leaves little margin before 2030. The argument lives or dies on whether the doubling pace persists and whether benchmark trends transfer to real work. Both premises are unsettled.
Benchmark tasks are cleaner and better specified than real software work. If benchmark horizon trends do not generalize to real-world work, and improvement plateaus or slows on higher-messiness tasks, then real-world horizons lag the benchmark curve, and month-long real-world tasks would not be automated by 2030.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 doubtful1 inconclusive
Granting that benchmark trends do not transfer and that progress slows on messy tasks, real-world month-long automation by 2030 becomes unlikely. It does not become impossible, because a faster underlying trend could absorb a fixed lag. The argument rests on a slowdown specific to messy tasks and on the denial of transfer from benchmarks. METR's own data give limited support to both, and neither is established.
Provenance
Where this claim has been said, linked to its canonical form.
If these results generalize to real-world software tasks, extrapolation of this trend predicts that within 5 years, AI systems will be capable of automating many software tasks that currently take humans a month.
Assessment history
0 status changes over 2 assessments. full history →
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.