The Race
Chapter 12 counted the fallen. Chapter 13 walked the moats. This chapter reports from the strip between them: capabilities where the human lead is real, measured, and shrinking on a clock anyone can read. Everything here is tagged [eroding] — which makes this the chapter most likely to be wrong by the time you read it. It says so up front, and it is built to be updated rather than believed: dates on every number, ranges instead of points, and a watch list at the end naming exactly which figures to check against the world.
The deeper purpose is the method. Numbers about AI arrive gift-wrapped in someone's excitement or someone's dread. This chapter's real lesson is how to unwrap them.
The clock [eroding]
The most useful single ruler for the race is METR's time-horizon metric, and it is worth understanding before quoting ⚙: take a large set of tasks whose difficulty is calibrated in human working time, and find the task length at which a model succeeds 50% of the time. That length is the model's horizon. It converts "how smart is it?" — unanswerable — into "how long a job can you hand it?" — measurable.
The readings, from the primary source. METR's original study (March 2025) found horizons doubling roughly every seven months. Its Time Horizon 1.1 update (January 29, 2026; 228 tasks) put Claude Opus 4.5's 50%-success horizon at 320 minutes — about 5.3 hours — with a confidence interval of 170 to 729 minutes, and GPT-5's at 214 minutes ⚖. The refit doubling time: 131 days measured from 2023, or 89 days measured from 2024.
Notice the honesty inside the primary source itself: the headline 5.3 hours arrives wrapped in a confidence interval running from under three hours to over twelve. That width is not sloppiness; it is what measuring a moving frontier with 228 tasks actually looks like. Distrust any secondhand version of this number that quotes the point and drops the interval ✦.
Sit with that pair of numbers a moment. The doubling time itself shrank between the two reports — the curve ran ahead of its own extrapolation within ten months. Whatever else is true, "the trend will simply slow down" has so far been the losing side of this particular bet.
One frontier reading, flagged at its true weight: secondary reports (a February 2026 data fit) place a March 2026 preview model near a 16-hour horizon — but that sits at the edge of what the task suite can measure at all, and METR itself warns that estimates above 16 hours are unreliable, since only 5 of the 228 tasks run that long ⚠ (~70% confidence). When a reading sits at the top of the ruler, suspect the ruler first.
Do the math, then doubt it
Here is the extrapolation everyone performs silently. Do it out loud instead, with both endpoints of the measured range. Anchor: 5.3 hours on January 29, 2026. Formula: horizon = 5.3 h × 2^(days elapsed ÷ doubling time). Working week = 40 hours; working month ≈ 170; working year ≈ 2,000.
| Date | At 131-day doubling | At 89-day doubling |
|---|---|---|
| January 2026 (measured) | 5.3 h | 5.3 h |
| January 2028 (+2 years) | ≈250 h — six working weeks | ≈1,600 h — nine working months |
| January 2031 (+5 years) | ≈80,000 h — a forty-year career | ≈8,000,000 h — four thousand working years |
Read the middle row first. Six working weeks is a serious consulting engagement; nine working months is a funded research project. Even the conservative endpoint, two years out, describes handing a machine the kind of work that structures a professional's whole quarter — at coin-flip reliability, on benchmark-clean tasks, if the trend holds. All three of those qualifiers are doing real work; keep them attached.
Read the bottom row again. Taken literally, it says a model in 2031 completes — at coin-flip reliability — a task that would occupy a human for several millennia. The two-year row is a forecast. The five-year row is a reminder that somewhere between the rows, the assumptions stop describing the world ✦. Exponentials are honest about the near term and hallucinate about the far term; that is not a flaw in this particular curve but a property of the method.
Four corrections before you quote even the near row:
- Reliability. These are 50%-success horizons — coin-flip odds. Horizons at 80% reliability run about four times shorter: 3–4 hours for the best models as of January 2026 ⚖. For "dependable," subtract roughly two doublings from everything in the table.
- The ruler's ceiling. The suite thins out above 16 hours (5 tasks of 228). Beyond that, the instrument — not necessarily the phenomenon — runs out ⚙. Watch for METR's successor suite before trusting any figure past that line.
- Benchmark tasks are cleaner than jobs. The suite's tasks are self-contained and machine-scored — Chapter 12's fallen profile, by design. Real months contain ambiguity, interruptions, and suppliers who lie. The next section prices that difference in dollars.
- Exponentials bend — in both directions. Every physical growth curve saturates eventually; this one, so far, has bent the wrong way for comfort (seven months to 89–131 days between March 2025 and January 2026). Anyone certain the curve must break, and anyone certain it must hold, is selling something ✦.
Now the arithmetic on the question people actually ask. From 5.3 hours to a 170-hour working month is five doublings, almost exactly. At 89 days per doubling, that lands around April 2027; at 131 days, around November 2027. Demand 80% reliability instead of 50% — two more doublings — and the window slides to late 2027 through mid-2028. Stack on the measurement caveats and the honest sentence reads: a dependable machine work-month is plausible somewhere in 2028–2029, if the trend holds — and "if the trend holds" is the entire bet. Both halves of that sentence were written in July 2026. Check the watch list before repeating either.
A year with consequences [eroding]
Against the optimism of the horizon curve, weigh the benchmark that made agents run a business.
Vending-Bench 2 (Andon Labs; updated July 2026) hands a model a simulated vending operation for a full business year: stock it, price it, negotiate with suppliers, handle the unexpected, stay solvent. No trick puzzles — just a simulated year of consequences. The best model, Claude Opus 4.7, finished the year with $10,937 in net worth. A merely "good" human strategy is estimated at about $63,000 ⚖. Roughly a six-fold gap — not on a logic exam, but on keeping a tiny business alive.
Where does the money go? Into precisely the gaps the horizon metric is too clean to see ⚙: reliability over long stretches, recovering from errors instead of compounding them, and not getting scammed by suppliers. Credit where due: the top models win points partly through persistent price negotiation — a socially fluent skill, and a small echo of Chapter 12's persuasion results. But fluency in the moment is not solvency over the year.
It is worth pausing on why a year of consequences resists what the theorem could not ⚙. Chapter 12's engine was the cheap verifier: try a million times, keep what the checker blesses. A business year has no checker mid-course — feedback arrives slowly, tangled with luck, and a wrong call in March surfaces as a hole in the accounts in September. Where feedback is slow, sparse, and consequential, brute search stops compounding, and the patient, error-correcting animal still earns its keep.
Two lines to hold together, both dated July 2026: models improve on this benchmark by about $799 per month, on a fit of R² = 0.96 — an unnervingly straight line — and that line crosses the $63,000 human mark around 2031 if nothing bends. So the same report card carries both messages at once: humans dominate the year-of-consequences today, and the gap is closing at a posted, public rate. That pairing — clear lead, visible clock — is what [eroding] means in this book.
The gap that embarrasses the machines [eroding]
The largest measured human–machine gap in mainstream benchmarking, as of July 2026, is not close.
ARC-AGI-3 (launched March 25, 2026) abandoned static grids for interactive environments with no instructions, no stated goals, no described rules. The agent must explore, build a model of the world from scratch, work out what the goal even is, and then plan toward it. It is a laboratory version of being dropped into an unfamiliar situation — the thing every puzzle in Part II rehearses in miniature.
At launch: humans solved 100% of the environments; every frontier model scored below 1% ⚖. Ordinary people, near-perfect; the best machines on earth, a rounding error above zero. By July 2026 the best public snapshot had reached about 8% (GPT-5.6 Sol; aggregator figure, ~80% confidence) ⚠.
Resist the urge to build a shrine here. This book's own sourcing gives the reason: ARC-AGI-1 stood for five years before falling inside a season (December 2024); ARC-AGI-2 stood for about sixteen months (July 2026). The family record argues the third one erodes too — and less-than-1% to roughly 8% in under four months is not a wall standing firm; it is a wall being climbed. Betting that this gap holds five years means betting against the exact trend that killed its two predecessors. Hence the tag: [eroding], with the largest current gap on the books. The ARC Prize 2026 season closes in November 2026; the watch list below states in advance what score would force a rewrite.
Insight, when the world talks back [eroding]
The same fault line runs through lateral thinking, and it is worth mapping precisely, because folk wisdom puts far too much of it on the human side.
On static riddle sets, the machines already do well: 76–83% across prompting styles on a 139-riddle benchmark (2026 study). The one-shot "aha" — the lightbulb item on a page — is no longer a reliable human lead, and this book will not pretend otherwise. Where humans still clearly win is interactive insight: situation puzzles (the parlor genre where a bizarre scenario must be cracked by asking the right yes/no questions) studied through LatEval and the ~975-puzzle SPLAT set. There, models show "almost negligible" lateral moves in interaction, defaulting to vertical, deductive grinding when the move that scores is to question the frame itself ⚖.
Note that this is the same failure ARC-AGI-3 measures, in different clothes ⚙: restructuring a representation while exploring it — updating not just your answer but your picture of what the question is, mid-conversation with a world that answers back. Part II called this breaking the frame; Chapter 4 trained it. As of July 2026, the interactive version remains human ground, sitting inside the blast radius of the fastest-moving benchmark in the field. [eroding] — enjoy it; train it; do not lean your identity on it.
The training implication is direct, and Part V takes it up: practice insight in its interactive form. Not the riddle on the page — the machines have that — but the version where you must buy information with questions, notice that your picture of the problem is wrong, and rebuild it while the clock runs. As of July 2026 that skill is measurably scarce in machines, and it is the most puzzle-like thing left on the board.
How to read a trend without hype
This section is the chapter's actual payload. The numbers above will age; the method for handling such numbers should not. Six rules, each one already exercised on this chapter's own data ✦:
- Date every number. An undated capability claim is advertising. Dated, it becomes a data point you can plot and check. Every figure in this Part carries its month for exactly this reason.
- Find the definition before the digits. Fifty percent of what? Preferred by whom? The ARC-AGI-2 "~85%" (July 2026) and "24.03%" (December 2025) describe the same benchmark under different compute budgets — a 3.5× disagreement produced entirely by the definition, not the capability.
- Extrapolate with both endpoints, never the midpoint. The measured doubling range, 89–131 days, diverges into a six-fold disagreement within two years (≈250 vs ≈1,600 hours by January 2028). An honest range that wide is not a weakness of the analysis; it is the analysis. A single confident number would be the lie.
- Ask what the ruler cannot see. METR's suite thins above 16 hours; leaderboards saturate; experts stop being able to tell (Chapter 13). A flat line at the top of a scale is a broken instrument as often as a plateau ⚙.
- Separate capability from consequence. A benchmark falling does not mean a job disappearing: the 67.1% diagnostic score arrived with over-ordered tests and no ward to answer to (Chapter 12); designed drugs had zero approvals as of mid-2026 (Chapter 12); the vending gap stood at roughly $52,000 (July 2026). Deployment runs on economics, workflow, and liability — Chapter 13's terrain — and lags capability by years.
- Precommit to what would change your mind. Decide today which number, at which threshold, flips your verdict — then check it on schedule. A forecast without a falsifier is a mood. The list below is this book's own precommitment, in print, where it cannot be quietly revised.
The watch list
Five numbers, checkable by anyone, that will grade this Part's homework. Status as of July 16, 2026:
| # | Watch | Where it stood at press time | What would move the verdict |
|---|---|---|---|
| 1 | IMO 2026, Shanghai, July 10–21 — live as this book went to press | prediction markets pricing an AI gold at ~82%; results due in the last week of July 2026 | gold: Chapter 12's "proof line" verdict confirmed on schedule · no gold: evidence of year-to-year variance the markets missed |
| 2 | ARC-AGI-3 leaderboard through the ARC Prize 2026 close (November 2026) | <1% at launch (March 2026) → ~8% (July 2026) | agents crossing 30–50%: downgrade "largest gap," shorten the interactive-insight half-life |
| 3 | METR's next time-horizon release (successor to TH 1.1, saturation addressed) | 5.3 h measured (January 2026); ~16 h preview unconfirmed ⚠ | confirmed >16 h horizons: the two-year column above is on track; slower: first real bend in the curve |
| 4 | Vending-Bench 2 versus the ~$63K "good human" line | $10,937 best; +$799/month, R² = 0.96 (July 2026) | a step-change (not slope): consequence-bearing agency eroding faster than linear; crossover ≈2031 if the line merely holds |
| 5 | Robot learning-from-experience (π*0.6 successors) | real-world RL demonstrated in narrow settings (2026) | robots that keep skills across days would attack two Chapter 13 moats at once — embodiment and continual learning |
A reader in 2028 holding this page has everything needed to grade it: five claims, five thresholds, five dates. Where the book turns out wrong, the method that wrote it shows how to correct it — which is the only kind of authority a chapter like this can honestly claim.
Where Part IV lands
Add up the three dispatches.
What fell was the closed loop: tasks that sit still, goals handed over complete, answers a machine can check for pennies — the proof, the contest, the exam, the prepared chart (Chapter 12). What held were the open loops: a world that pushes back, learning that compounds across decades, doubt that is owned rather than performed, goals that are chosen and signed for, the tails rather than the center (Chapter 13). And between them runs this chapter's front line — hours doubling every 89–131 days, a business year still six-fold human, an interactive-reasoning gap of 100% against 8% — every figure dated, every figure moving (as of July 2026).
Now the moral, and it is the reason this Part sits inside a puzzle book rather than a policy report.
Look at the durable column one more time. Learning deeply from a handful of examples. Carrying a lesson forward so it compounds. Knowing which of your own answers to doubt. Restructuring the problem while you are inside it. Choosing what deserves the work, and signing for the result. That is not a list of consolations — it is a curriculum, and it is this book's curriculum. One puzzle digested into a permanent Move is Moat 2 in action. The mistake journal and upsolving are Moat 3, practiced. Calibration drills are Moat 4. The setter's mind — asking why this problem, why these constraints — is Moat 5. And every frame this book has broken since the nine dots is the muscle ARC-AGI-3 measures.
None of this was designed as a response to the machines — the canon is millennia older than the transistor. But the machines have, without meaning to, run the audit: of everything the gymnasium teaches, they took the checkable and left the compounding.
The gymnasium is four thousand years old. The machines did not close it; they surveyed it, and marked which halls still matter. The scoreboard will keep moving — the dates are on every number so you can move with it. Part V turns the map into a training plan.
Sources
All accessed July 16, 2026. Press and aggregator items are marked; treat their figures with the confidence noted in the text.
- METR, Time Horizon 1.1 (January 29, 2026): metr.org/blog/2026-1-29-time-horizon-1-1/
- METR, Measuring AI Ability to Complete Long Tasks (March 19, 2025): metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
- Doubling-time refit and ~16 h preview figure (February 2026 data fit) — secondary: x.com/aakashgupta/status/2053207639107174452
- Vending-Bench 2, Andon Labs (updated July 2026): andonlabs.com/evals/vending-bench-2
- ARC-AGI-3, ARC Prize Foundation (launched March 25, 2026): arcprize.org/arc-agi/3 · technical report: arxiv.org/abs/2603.24621
- ARC-AGI-3 frontier snapshot ~8% (July 2026) — aggregator figure, ~80% confidence; no primary publication as of press time
- ARC Prize 2025 results and analysis (efficiency-track figures, December 2025): arcprize.org/blog/arc-prize-2025-results-analysis · arxiv.org/html/2601.10904v1
- SPLAT situation-puzzle benchmark (~975 puzzles): arxiv.org/html/2410.06733v1 · LatEval — cited by name
- Riddle-benchmark results, 76–83% across prompting styles, 139 riddles (2026 study) — cited by description; no public URL recorded at press time
- International Mathematical Olympiad 2026, Shanghai (July 10–21, 2026): imo-official.org/editions/2026/
- Chapter cross-references: diagnosis and drug-approval status (Chapter 12 sources); expert-detection flip and accountability doctrine (Chapter 13 sources)
As of September 2026
Added after press time; the July text above is unrevised. Checked September 2, 2026.
- Watch #1 resolved — gold, and past it. At IMO 2026 (Shanghai, July 10–21), RedNote's dots-note-3.0 scored a perfect 42/42, the first LLM to do so, and GPT-5.6 Pro also solved all six problems ⚖. Chapter 12's "proof line" verdict is confirmed on schedule.
- Watch #2 tripped — ARC-AGI-3. Claude Opus 5 scored 30.2% on July 24, 2026, roughly quadrupling the ~8% above and crossing this chapter's own 30–50% falsifier ⚖. On August 24, 2026, NVIDIA's AVO agent — Opus 5 inside a long-horizon scaffold — cleared all 183 levels of the 25-environment public set at 100%; the semi-private and private sets remain untested ⚠.
- What that costs the July text. "The largest measured human–machine gap in mainstream benchmarking" no longer describes ARC-AGI-3, and "100% against 8%" is a July reading, not a current one. The [eroding] tag held; the half-life was shorter than this chapter estimated — five months from launch, not the sixteen its predecessor lasted.
- Unmoved. No Vending-Bench 2 or METR time-horizon release dated after July 2026 was found; watch items #3, #4 and #5 stand where the table left them.
- ARC Prize, Claude Opus 5 results: arcprize.org/results/anthropic-claude-opus-5 · NVIDIA AVO (August 2026): developer.nvidia.com/blog/nvidia-avo-reaches-100-on-arc-agi-3-demonstrating-a-frontier-level-general-purpose-architecture-for-long-horizon-autonomous-agents/ · IMO 2026: imo-official.org/editions/2026/