Calendar year when an AI company first internally develops an AI that can do any coding task its best engineer can do, faster and cheaper
- Measures
- The calendar year in which any AI company first develops internally (whether or not deployed publicly) an AI system able to perform any coding task its best engineer can perform, while being faster and cheaper than that engineer: the "superhuman coder" milestone as defined in AI 2027 and close to the AI Futures Project's later "Automated Coder".
- Population
- Frontier AI developers worldwide
- Conditions
- Dated by when the capability first exists inside a company, as later established; the milestone may never occur.
- Unit
- a calendar year
Most likely about 2030; 80% interval 2027 to 2038, with a 7% chance that The milestone is not reached by 2050 or ever (e.g. a capability plateau, regulation, or catastrophe halting development).
Stated as a distribution, the precision the evidence supports. Right-skewed with a heavy tail into the 2030s and beyond.
Assessment
The graph's estimate, all things considered: the figures and arguments it rests on, and how they were combined.
The "superhuman coder" milestone, an AI that can do any coding task a company's best engineer can do while being faster and cheaper, was placed in 2027 by the April 2025 AI 2027 scenario, whose forecasters all named 2027 as one of the most likely years. By November 2026 it has not been reached, and the scenario's own authors have since lengthened and then partly shortened their forecasts: their April 2026 update put Daniel Kokotajlo's median for the closely related "Automated Coder" milestone at mid-2028 and Eli Lifland's at mid-2030, and their August 2026 update moved slightly shorter again while judging that progress was running at about 70 to 90 percent of AI 2027's pace.
These forecasts rest on extrapolating rapidly growing METR time horizons, reported coding uplift inside AI companies, and AI-company revenue. Each extrapolation carries real uncertainty about what level corresponds to full automation, about whether benchmark trends generalize to real-world engineering, and about whether compute growth continues. Forecasters outside this group generally place the milestone later.
The estimate here has a median of 2030, with a 90 percent interval running roughly from 2027 to the late 2030s, and about a 7 percent chance that the milestone is never reached on any relevant horizon. 2027 remains plausible but is now less likely than not. One further caution: the quantity dates internal development, which may not become public right away, so the year may only be established afterward.
The work: which figures, how they were combined, and the arithmetic
Starting point: the AI 2027 instance (April 2025) states 2027 as one of the most likely years; its timelines supplement gave within-model medians of 2027 to 2028, while Kokotajlo's own median was later said to be about 2028 and Lifland's about 2030 to 2031 (blog.aifutures.org, "Clarifying how our AI timelines forecasts have changed"). The December 2025 AI Futures Model said the new model with median parameters predicts about 2032 for SC (blog.aifutures.org/p/ai-futures-model-dec-2025-update). The Q1 2026 update (2 April 2026, as reported by the ai2027-tracker changelog) moved Kokotajlo's Automated Coder median from late 2029 to mid-2028 and Lifland's from early 2032 to mid-2030. The Q2.5 update (16 Aug 2026, blog.aifutures.org/p/q25-2026-timelines-update-uplift) says timelines got slightly shorter. It adds coding-uplift and revenue anchors, notes Anthropic internal survey uplift rising from 1.25x to 4x in seven months, and judges reality at 70 to 90 percent of AI 2027's pace (which at 75 percent would give AC in mid-2027). These forecasts are explicitly conditional on progress going as fast as technically feasible, so unconditional dates should be somewhat later.
Definitional note: AI 2027's SC (30x agents at 30x speed on 5% of compute) is somewhat stronger than AIFM's Automated Coder (a company would rather fire its engineers than forgo AI coding). This quantity's wording ("any coding task... faster and cheaper") sits between them, closer to SC, so I shift a few months later than AC medians.
Combination: authors' medians of mid-2028 to about 2030 for AC, plus a shift for the SC/AC gap, the conditional-on-feasibility caveat, and wider expert opinion (most forecasters outside the AI 2027 group, and the Metaculus/LEAP-style forecaster pools, have later dates). Together these give a median of about 2030. The lower tail is truncated by non-attainment through November 2026, so 2026 itself is near zero; about 10 percent by end-2027 reflects fast uplift trends. The upper tail is heavy for benchmark-to-reality gaps, compute slowdowns, and plateaus. About 7 percent is assigned to never, or not before 2050.
Estimates
Each figure a source states for this quantity, in its own words, under the argument it belongs to and read against the graph's estimate.
Grey band: the graph's 90% interval, with its central value as a line. Dark marks: each source's own figure. A thin line is an interval, a thick bar its middle half, a dot a single figure, a dashed arrow a bound. An open box is where an argument places the value.
Decomposition
The arguments the estimate rests on, each stated as it runs, with what it shows about the value and its premises linked inline. ↗︎ opens a premise; the map shows how they fit together.
Because the length of software tasks frontier AI agents can complete has grown roughly exponentially, and coding uplift and AI-company revenue have grown similarly fast, extrapolating these trends to the level at which a company would rather replace its engineers than forgo AI places the milestone around 2027 to 2032. The extrapolation presumes that benchmark-measured capability trends generalize to real-world software engineering and that the trends continue without a compute or data slowdown.
- 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 inconclusive1 not separately assessed
If its premises hold, trend extrapolation makes a late-2020s to early-2030s arrival much more likely, though it cannot pin down a single year: the level that corresponds to full automation is uncertain, and so is whether the trends keep their shape. The argument depends on continued rapid growth in time horizons and, above all, on the framework premise that benchmark trends generalize to real-world engineering. If that premise fails, every date the argument yields is too early.
Provenance
Where this quantity's evidence comes from. The figures sources state are under Estimates above.
All forecasters place 2027 as one of the most likely years in which an SC might be developed
An SC (superhuman coder) is defined as an AI system that can do any coding tasks that the best AGI company engineer does, while being much faster and cheaper.
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.
Created by extractor · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.