A Forecast with a Clock
On 20 January 2026, at the World Economic Forum in Davos, Anthropic CEO Dario Amodei said the company's models might be six to twelve months from doing “most, maybe all” of what software engineers do, end to end. He added that some engineers at Anthropic had stopped writing code by hand and instead edited what the model produced.
Amodei also named limits on the pace: chip manufacturing and model-training time. He runs a company that sells these systems, so the forecast comes from an interested party. It is nevertheless specific enough to test.
Six to twelve months from 20 January creates a window from July 2026 to January 2027. Its earliest date has passed. Its latest is three months away. Nine months of evidence can now be compared with the claim.
What Has Changed in Capability
The most cited yardstick is METR's “time horizon”: the length of a task, measured in human-expert hours, that a model completes successfully half the time.
On METR's updated suite, Claude Opus 4.5 sat at roughly five hours at the start of the year. Opus 4.6, released in February, was estimated at around 12 hours, with an interval extending from approximately five to 66 hours. In March, METR evaluated an early version of Claude Mythos Preview and placed its horizon at no less than 16 hours, with a 95% interval of 8.5 to 55 hours.
Three caveats matter. METR says measurements above 16 hours are unreliable on its current suite, which contains only five tasks that long. The tasks are clean and automatically scored, unlike most commercial software. At an 80% success rate, the same Mythos evaluation falls to about three hours—the more relevant figure where one failure in five is expensive.
Even with those limits, capability has grown faster than many expected. Forecaster Ajeya Cotra has publicly said she underestimated it again.
What Has Not Changed in the Job Market
The Yale Budget Lab's ongoing tracker has not found a clear AI-related footprint in economy-wide employment data.
Indeed Hiring Lab reports US software-development postings up almost 15% since late February 2025, while postings overall fell 7%. The rebound begins from a low base: software-development postings remain approximately 27.5% below their pre-pandemic level.
The US Bureau of Labor Statistics projects 10% employment growth for software developers, quality-assurance analysts and testers between 2025 and 2035, with about 106,100 openings per year. That is lower than the 15% projected in the previous 2024–2034 edition, but it is not a forecast of occupational collapse.
What Has Changed: the Bottom Rung
Stanford Digital Economy Lab, using ADP payroll data through June 2026, finds employment among 22-to-25-year-olds in highly AI-exposed occupations approximately 19% below where it would be had it kept pace with peers in less-exposed work. A year earlier, the gap was 15%.
The difference is driven mainly by reduced hiring rather than layoffs. Experienced workers in the same occupations show no comparable gap. The authors describe the results as early, descriptive indicators rather than causal estimates, and the figure covers all highly exposed occupations, of which software development is one.
Indeed's posting data shows the same split. In the first quarter of 2026, senior roles represented 69.3% of software-development postings and entry-level roles only 4.5%—the lowest entry-level share of any sector Indeed tracks. Senior positions accounted for 71% of the growth in postings between May 2025 and May 2026.
Are Developers Actually Faster?
The evidence is thinner than the enthusiasm. METR's 2025 randomised trial found that experienced open-source developers took approximately 19% longer when using AI tools, even while believing they were faster. METR has since marked that result as out of date.
Its 2026 follow-up produced an unreliable signal because a growing number of developers declined to participate rather than work without AI. METR considers it likely that developers are now more accelerated than they were a year ago, but says its data provides only very weak evidence about the size of the effect.
No clean, current measurement yet establishes how much faster professional software teams have become across realistic production work.
What the Evidence Adds Up To
This section is analysis, not established fact.
Writing code is becoming cheap. Verifying it is not. The result of hours of autonomous agent work is still a change that someone must read, test and answer for, especially when it touches payments, personal data or security. Indeed's economists interpret the postings similarly: demand is growing for experienced people who can work with AI.
That creates a problem the headline debate overlooks. Senior engineers are made by doing junior work. If that work disappears, the supply of future seniors contracts, and the shortage of reviewers may become most acute just as demand for them peaks.
The second consequence concerns business. When a working product can be assembled over a weekend, merely having one stops being an advantage. Customer knowledge, distribution, trust and support decide who wins. A founder's question changes from “Can we build it?” to “Does anyone want it, and how will they find us?”
What Businesses Should Do
- Measure outcomes, not output. Track delivery time, defects and incidents rather than volume of code or number of pull requests.
- Keep review where errors are expensive. Human accountability remains essential for actions that are costly or irreversible.
- Validate demand before building. Cheaper prototypes make validation easier, not unnecessary.
- Protect the learning pipeline. If you hire, decide deliberately how junior engineers will acquire the judgment expected of future senior staff.
Verdict So Far
Read literally—“most, maybe all, end to end”—the prediction has not come true. Employment data does not show it, and no current evaluation shows models reliably owning complex software work for months.
But the capability half of the forecast looks closer than it did in January, and the labour-market damage is real, although narrow. The earliest date in Amodei's window has passed without confirmation. The latest is three months away.
The test is not simply whether engineers disappear. It is whether the next generation can still get started.
Sources
- World Economic Forum, “The Day After AGI” session, Davos, 20 January 2026 — Radio Davos episode.
- METR, Time Horizon 1.1, 29 January 2026.
- METR, evaluation of an early Claude Mythos Preview snapshot, March 2026.
- METR, developer productivity experiment update, February 2026.
- Ajeya Cotra, I Underestimated AI Capabilities (Again).
- Stanford Digital Economy Lab, Canaries in the Coal Mine?, revised 12 August 2026.
- Indeed Hiring Lab, AI and Job Postings: From Destruction to Creation?, 8 July 2026.
- Indeed Hiring Lab, The Labor Market Is Tilting Toward Seniority, 23 July 2026.
- Yale Budget Lab, Tracking the Impact of AI on the Labor Market.
- US Bureau of Labor Statistics, Occupational Outlook Handbook: Software Developers.
Write a comment