Direct answer: AI will not become cheaper than human labor everywhere on one date. The cost of three different things (a token, a finished task, an accountable outcome) crosses human cost at different times, and by my calculation the middle one lands between mid-2027 and late 2029 depending on the reliability you demand. The first effects are already measurable, and they appear as missing entry-level hires, not layoffs.
The comparison everyone makes is the wrong one
A million tokens costs less than a few minutes of an analyst’s time. Screenshot that, add a hockey-stick chart, and you have the standard “AI is cheaper than people” post.
It proves nothing, because nobody buys tokens. They buy finished work. A finished task includes the retries, the review, the cleanup of the 20% that came back wrong, and someone to blame when it ships broken. Compare that to a salary and the answer changes with every variable in the sentence.
The “2028” in the title has a real origin. METR’s 2025 time-horizon paper extrapolated a seven-month doubling and projected AI agents reaching tasks of about one work-month (167 hours) between late 2028 and early 2031 (as summarized in a critical review; I did not open the original PDF). The date was never a cost forecast. It was a capability extrapolation, and cost has to be layered on top.
Three crossings, not one
The diagram is only a map. The mechanism sits in crossing 2, so that’s where the numbers go.
Crossing 2: the calculation
METR’s task-completion time horizon measures how long a task (in human working time) an AI can finish at a given success rate. Its January 2026 update (Time Horizon 1.1, cited via Wikipedia, a secondary source) puts the post-2023 doubling time at about 131 days, against roughly seven months for 2019–2024. As of May 2026, the top model, Claude Mythos, sits at a 50% horizon of at least 16 hours and an 80% horizon of 3 hours 6 minutes. A forecasting page puts the headline figure near 17.4 hours and says the series is pinned at the instrument’s ceiling as of late August 2026. METR itself reportedly warns that estimates above 16 hours are unreliable because the task suite is saturating, so treat 16 hours as a floor.
Here is what those inputs imply. This is my arithmetic, not a forecast:
- Starting point: 16 h (50%) and 3.1 h (80%), as of 9 May 2026.
- Target: one work-month, 167 h.
- Doubling time: 212 days (the older 7-month fit) or 130.8 days (the post-2023 fit).
| Horizon reaches 167 hours at… | Slow fit (212 d) | Fast fit (130.8 d) |
|---|---|---|
| 50% success rate | ~April 2028 | ~July 2027 |
| 80% success rate | ~September 2029 | ~May 2028 |
Demanding 80% instead of 50% reliability pushes the date out by 10 to 17 months. The reliability threshold is a bigger lever than the choice between the two doubling fits. The “2028” in your headline sits inside this band, not at a point on it.
What this table does not tell you: horizons measure software-style tasks with checkable outputs. The ratio between 50% and 80% horizons may not hold. And nothing here includes cost.
What it costs, and why the price curve misleads
Epoch AI’s analysis of inference prices found that the price to reach a fixed benchmark level fell at a median of 50x per year, with a range of 9x to 900x. Epoch itself notes that the fastest declines occurred in the most recent window, so it is unclear they will persist. A separate paper (summary here) estimates 5–10x per year for frontier models, and about 3x per year for pure algorithmic efficiency once open models and hardware prices are controlled for.
These numbers disagree because they measure different things: price at a fixed capability level versus price of the frontier. Which one you extrapolate decides whether the crossing is next year or in five.
Here is the non-obvious part. Epoch’s trends dashboard puts chip performance per dollar at about +37% per year. Multiply that by the ~3x algorithmic gain, and you get roughly 4x per year of real cost reduction (3 × 1.37 ≈ 4.1). Observed prices fell much faster than that. My inference, not a sourced finding: the gap is competition, margin compression, and distillation, and those have a floor. The curve most people extrapolate is partly a price war. The curve that governs 2028 is the 4x one.
Does cost per task explode as capability grows? Two analyses disagree, and both are single-author work built on METR data:
- Toby Ord’s reading of METR’s charts (2025-era models) found AI “sweet spots” of 40¢ to $40 per hour against a best-case human software engineer at about $120, but o3 at the edge of its horizon cost about $350 per hour, above the human. Cost balloons at the limit of capability.
- A Redwood Research analysis estimates token cost from OpenRouter pricing and finds the cost ratio at the 50% horizon isn’t rising. Capping spend at 1/32 of human cost barely slows the horizon trend.
Ord’s data is older and measures the edge; Redwood’s is newer and measures the typical trajectory. They are not measuring the same point on the curve, which is why they don’t conflict. I weigh Redwood more for tasks well inside a model’s horizon and Ord more for tasks at its limit.
METR’s July 2026 expenditure horizon is the first attempt to measure the dollar budget at which human effort starts beating agent effort. The proof of concept uses a single benchmark (the NanoGPT speedrun), and, per press coverage, human cost came from interviewing two contributors plus an AI model’s estimate. Useful method, weak evidence.
My position: for tasks comfortably inside the 80% horizon, price is no longer the binding constraint. Reliability and verification cost are. The sensible test is the one below.
A formula worth running on your own work
Expected cost per accepted output = C_ai + (1 − p) × C_fix
Where C_ai is the agent run cost, p is the pass rate, and C_fix is what a failure costs you in review and rework. Illustrative inputs, not data: a human does an 8-hour task for $480. The agent run costs $40.
- p = 0.8, C_fix = $240 → $88. AI wins easily.
- p = 0.5, C_fix = $600 (long tasks fail expensively) → $340. Still cheaper, but the margin has thinned.
- p = 0.2, C_fix = $600 → $520. Human wins.
The crossing is driven by p and C_fix. The token price is nearly irrelevant.
Where the shock lands first: the hiring margin
Stanford’s Digital Economy Lab (revised August 2026) uses ADP payroll data and finds no economy-wide displacement. But employment of workers aged 22–25 in highly AI-exposed occupations stands 19% below where it would be had it tracked less-exposed peers, with no comparable gap for experienced workers. The adjustment runs mainly through reduced hiring, not separations. Where AI complements workers, employment is flat or rising. Pay has not moved much so far.
One caution on the numbers: the figure has drifted across versions. The paper’s earlier headline was 13%, a March 2026 summary cites 16%, and the revised paper reports 15% at the July 2025 data vintage and 19% as of June 2026. Different data cuts and methods, so the trend (widening) is better evidence than any single number. The authors themselves say no single study is definitive.
The Yale Budget Lab reaches a milder reading: the US occupational mix has not shifted faster than in earlier technology waves, and the 20–24 cohort’s slightly larger shift may reflect general labor-market weakness. An EPI analysis of the class of 2026 likewise finds no clear collapse in tech-sector employment for young graduates.
These studies are not equally strong on the same question. Yale’s instrument measures the whole workforce’s occupational mix. A cohort that is a small share of total employment barely moves it. Stanford’s design isolates exactly that cohort. For “is the entry rung being removed?”, I rank Stanford’s payroll evidence highest. For “is the whole economy being reshuffled?”, Yale’s null is the better-matched evidence, and the answer so far is no.
The shock looks like a job posting that never appears. It doesn’t show up in unemployment claims.
Why “everywhere” fails, and where it fails first
AI costs the same in Manila and Milwaukee. Human wages don’t. So you’d expect the crossing to arrive first where labor is dearest. The evidence points to a different first mover: work that is already traded across borders at a price.
The World Bank’s World Development Report 2026 puts high automation risk at 4.5% of jobs in low- and middle-income countries versus 14.2% in high-income ones, yet names the Philippines among the most exposed economies, because its export earnings depend on customer support, transcription, and document processing. Coverage of the report also cites a 39% decline in outsourced jobs flowing to developing economies in 2025, a figure I could only verify through that secondary report. The Philippine industry body IBPAP reportedly cut its forecasts in July 2026, again via a secondary summary.
My reading: offshore labor is already bought by price comparison, with contracts, SLAs and per-seat rates. A buyer who already compares vendors on cost will compare an AI vendor the same way. Domestic work wrapped in employment norms, licensing, and local context has more friction. The first “everywhere” crossing is therefore a crossing of procurement, not of wages.
The open questions
What happens to wages? I can’t answer this yet. Stanford sees adjustment in headcount, not base pay. If AI sets a price ceiling for codifiable tasks, you’d expect wage compression at the entry level to follow once hiring has already shrunk. Resolving it takes cohort-level payroll data on new-hire starting pay by exposure, over at least four more quarters. Stanford’s monthly Canaries dashboard is the right instrument.
Does compute supply cap any of this? I did not research supply constraints (energy, chips) for this piece, so I make no claim.
What to do with this
- Stop benchmarking tokens against salaries. Log cost per accepted output for three task types you actually run, with reviewer minutes included. If you can’t measure p and C_fix, you can’t know which side of the line you’re on.
- Demand 80%, not 50%, when you size a deployment. The table shows that choice moves the date by up to 17 months.
- If you buy outsourced services, re-tender on per-resolution or per-accepted-output pricing, not per-seat. Seat pricing leaves all the AI savings with the vendor.
- If you hire junior staff, the evidence says the risk sits in routine, codified tasks, where AI substitutes for rather than assists the worker. Design entry roles around verification and exception handling, which is where complement-type usage shows stable employment.
- Track four signals through 2028: METR’s 80% horizon on a non-saturated suite; Stanford’s monthly Canaries update; price at fixed capability (Epoch); and Philippine/Indian BPO headcount versus revenue.
The shock is not a date on a calendar. It is a set of thresholds, and each one is crossed by whoever measures their own cost per accepted output first.