The Verification Frontier: what actually limits the next decade of AI

Look at what is actually being scaled. Training compute for notable models has grown 4.5× per year since 2010. Algorithms improve on top of that: each year the same performance takes 3× less compute. The installed stock of AI chips has grown 3.4× per year since 2022, doubling roughly every 7 months. Context windows have grown 30× per year since 2023 — a doubling every 2.4 months.^[1] On the demand side, the price of a fixed capability collapses: across six benchmarks Epoch AI measures annual price declines from 9× at the slowest to 900× at the fastest, median 50×, rising to a median of 200× when the sample is restricted to models released after January 2024.^[2]

Every one of those curves is a curve about producing output. None of them is a curve about knowing whether the output is any good. And the second question is the one that gates deployment, gates training, and gates trust.

This is not a philosophical complaint. It shows up as a hard number in the one place the industry has agreed to measure long-horizon competence. METR's time horizon is the length of task — measured in how long it takes a human expert — that an agent completes at a given reliability. On software tasks the 50% horizon has been doubling every ~7 months for six years, possibly accelerating to every 4 months in 2024.^[3][4] But reliability is the whole story: current models are near 100% on tasks under 4 minutes and succeed under 10% of the time past about 4 hours.^[3] The failure is not that the model cannot do the work. It is that nobody, including the model, can tell which of the four-hour attempts worked.

The idea

Treat verification as the priced input it is. For any unit of useful work, write down what it costs to make a candidate and what it costs to confirm one. The ratio between those two numbers predicts, better than parameter count or FLOP budget, how fast that domain will move — and it explains why the fields that look like they are exploding and the fields that look stuck are being driven by the same models in the same year.

Sort the domains they measured by how expensive it is to establish ground truth, and the doubling times sort with them.

At the cheap end sit math contests, competitive programming, scientific QA and semi-realistic software engineering. Every one of those has a mechanical oracle — an exact answer, a test suite, a scorer you can run ten thousand times for the price of electricity. METR finds these cluster at 50–200+ minute horizons, doubling every 2–6 months; Mock AIME comes in around ~3 months, and METR's own software suite around ~4 months.^[4]

At the expensive end sits everything that has to be checked against a world. Visual computer use — OSWorld, WebArena — has horizons 40–100× shorter than that cluster. Tesla's self-driving is improving at about 0.6 doublings per year, a doubling time near 20 months, which is five to ten times slower than the verifiable cluster.^[4] The models driving all of these are the same generation of models. What differs is what it costs to find out whether an attempt succeeded.

The Stanford AI Index makes the same point from the benchmark side without framing it that way. In one year, SWE-bench Verified went from 60% to near 100% — a benchmark whose entire premise is that each task ships with a test suite that decides the matter. In the same report, the top model reads an analog clock with 50.1% accuracy, and agents on structured benchmarks still fail roughly one task in three.^[6] Saturation arrived precisely where the verifier was free.

Why this is not just a restatement of "some tasks are harder"

Difficulty and verifiability come apart, which is what makes the claim non-trivial. Competition mathematics is genuinely hard — models now meet or exceed human baselines on it^[6] — and it moves fast because a correct answer is a string you can compare. Reading a clock face is trivially easy for a person and moves slowly, because there is no corpus of clock-reading tasks with machine-checkable labels and no reason anyone built one. Hardness predicts the current level. Verifiability predicts the slope.

The stronger version of the claim: the data wall is a verification wall. Epoch AI puts the effective stock of human-generated public text at about 300 trillion tokens (90% CI 100T–1000T), fully used somewhere between 2026 and 2032 at 80% confidence.^[5] Synthetic text is unlimited and costs nearly nothing — the constraint was never volume. What runs out is text carrying a signal somebody paid to establish. You can generate a trillion tokens of plausible reasoning tonight; you cannot generate a trillion tokens of confirmed reasoning without a trillion confirmations.

Fix a task. Let \(g\) be the cost of generating one candidate solution, \(v\) the cost of verifying one, and \(p\) the probability that a candidate is correct. Under the sample-and-check loop that every modern system runs — best-of-\(n\), rejection sampling, agentic retry — the expected cost of obtaining one trusted solution is

\[ C \;=\; \frac{g + v}{p}. \]

Define the verification ratio \(\rho = v/g\), so that

\[ C \;=\; \frac{g\,(1 + \rho)}{p}. \]

Everything the industry is scaling drives \(g\) down: cheaper tokens, faster chips, better algorithms. But \(C\) is bounded below by the verification term. As \(g \to 0\) with \(v\) fixed, \(C \to v/p\), and further progress on generation buys nothing at all. The entire cost of the enterprise migrates into \(v\) — and once \(\rho \gg 1\), the observed doubling time in a domain is a fact about its verifier, not about its models. That is the formal content of §02.

Why executable checkers win by so much

Verifier cost has a fixed and a marginal part. Writing the checker costs \(V_0\) once; each use costs \(c\). Amortised over \(N\) uses,

\[ v \;=\; \frac{V_0}{N} + c. \]

A test suite has a real \(V_0\) and a marginal \(c\) close to zero, and \(N\) is unbounded — so \(v \to 0\) and the domain enters the fast regime permanently. Human review has \(V_0 \approx 0\) and an irreducible \(c\): it never amortises, no matter how many times you do it. This is the whole gap. It is not that software engineering is easier than radiology; it is that software engineering's verifier was written once in 1998 and has been running for free ever since.

A ladder of verifiers

Ordered by trustworthiness per unit cost — and, empirically, by how fast the domains sitting on each rung are moving:

• Formal. A proof kernel or type checker. Verdict is sound by construction, \(c \approx 0\). The narrowest coverage and the highest \(V_0\).
• Executable. Tests, compilers, simulators, exact-answer keys. Very cheap per use, high coverage in code and mathematics. This is the rung the last three years of progress were built on.
• Empirical. A physical trial, an A/B test, a market that settles. Trustworthy and irreducibly slow — \(c\) is measured in days and dollars, not milliseconds.
• Human. Expert review. Broad coverage, never amortises, and the bottleneck scales with the number of experts rather than the number of chips.
• Learned. A reward model or a model asked to grade another model. Cheap and general, and the only rung whose verdicts are themselves unverified. §06.

The framework's prediction is uncomfortable and reasonably falsifiable: the next decade's returns accrue to whoever moves capabilities down this ladder — turning empirical and human checks into executable ones. Not to whoever trains the largest model.

Escape 1 — learn the verifier

The obvious move is to train a model to do the checking: a reward model, or simply a stronger model asked to grade a weaker one. This is cheap, general, and it is what reinforcement learning from verifiable rewards (RLVR) industrialised. It also relocates the problem rather than dissolving it, because a learned verifier's verdicts are themselves unverified — you have bought coverage by giving up soundness.

The measured cost of that trade is specific. Wu et al. find that RLVR reliably enhances precision and consistently improves pass@1, but that it may progressively narrow exploration and potentially overlook correct yet underrepresented solutions. Under larger sampling budgets, the shrinkage of empirical support generally outweighs the expansion — the trained model fails to recover correct answers the base model could already reach. Token-level entropy sometimes rises while answer-level entropy declines: more apparent uncertainty per step, converging on a smaller set of distinct answers.^[8] Optimising hard against a checkable signal makes a model better at the checked thing and narrower everywhere else. That is not a route around the verifier. It is the verifier's shape being pressed into the model.

Escape 2 — spend test-time compute instead

The second move is to let the model deliberate longer, on the theory that enough internal search substitutes for external checking. Where a mechanical oracle exists inside the loop, this works well — it is most of why reasoning models advanced the Epoch Capabilities Index at 14 points a year rather than 6.^[1]

Where no oracle exists, it stops working, and the failure is clean. Zhao, Hooi and Ng evaluate 14 reasoning models across three knowledge-intensive benchmarks under reasoning-effort, thinking-budget and budget-forcing controls. Their finding: increasing the reasoning length results in minimal or no accuracy gains across most models and all benchmarks. o4-mini shows more than 8 times increase in reasoning tokens with accuracy almost unchanged on all tasks; on FRAMES, GPT-5 and gpt-oss-20b gain 3–5 percentage points for reasoning lengths scaling over 8×; one setting consumes nearly 20× more reasoning tokens.^[9] You cannot verify a fact by thinking about it harder. Test-time compute is a search algorithm, and a search algorithm without a scoring function is just a longer walk.

Escape 3 — deny the premise

The strongest objection is that capability is plainly accelerating and this whole frame is pessimism looking for a mechanism. The AI Index supports that reading on its face: SWE-bench Verified from 60% to near 100% in a single year, human baselines exceeded on several hard evaluations, organisational adoption at 88%, $285.9 billion of US private AI investment in 2025.^[6]

The response is that the accelerating results are, almost without exception, the verifiable ones — and the same report says evaluation is now the thing falling behind. If the fastest-moving numbers are the ones with the cheapest oracles, "capability is accelerating" and "verification is the constraint" are not competing claims. They are the same observation from two ends.

Where this report is weakest

Stated plainly, because the framework deserves the same scrutiny as the escapes:

• It is observational. Nothing here isolates verifier cost experimentally. The domain comparison in Table 1 is a correlation across benchmark families that differ in many other ways.
• The causal arrow may run backwards. Verifiable domains attract more benchmarks, more RL effort and more training data because they are verifiable. Cheap verification may be the consequence of investment rather than the cause of progress, and the published record cannot currently separate the two.
• The evidence base is thin. Table 1 rests on one research group's methodology applied to about ten benchmark families. METR itself does not report a 99% time horizon because it would require many more tasks.^[10]
• The verifier-rung column is an interpretation. No source classifies domains this way. Reasonable people would sort some of those rows differently.
• \(g\), \(v\) and \(p\) are never jointly measured. The model in §03 is not fitted to anything. Its value is organisational, and it should be treated as a hypothesis with a clear falsification: find a domain with an expensive verifier that is nonetheless doubling in months, and the claim fails.

What survives all of that is modest and, I think, hard to dispute: generation cost has fallen by two to three orders of magnitude a year while no comparable curve exists for verification cost, and nobody has published one. Until somebody measures \(v\) the way Epoch measures \(g\), the most important number in the field is the one nobody is tracking.