Models double their attention span every four months. Are you doubling your demonstrations?

METR's latest data shows AI task horizons accelerating faster than anyone projected. The bottleneck isn't compute: it's the human demonstration data that teaches models to sustain focus across real workflows.

In January 2026, METR released Time Horizon 1.1, an update to the benchmark that tracks how long AI models can work autonomously before their success rate drops below 50%. The headline number: the time horizon at which frontier models hit 50% success has been doubling every 131 days since 2023. Zoom into the post-2024 data alone, and the doubling time compresses to 89 days. Claude Opus 4.5, the current frontier, holds a 320-minute time horizon, over five hours of sustained autonomous work.

Extrapolate that curve and models reach a full workday by 2027, a full work-week by 2028. That trajectory is the measured trend across 228 tasks and 14 frontier models. And it creates a problem almost no one in healthcare or banking is talking about: the models are ready for longer tasks, but the training data to teach them barely exists.

The trendline has a data problem

METR's updated benchmark doubled its inventory of long tasks, those estimated to take a human eight or more hours, from 14 to 31. That expansion matters because earlier versions were saturating at the top end. Models were running out of hard problems to fail on, making the trendline look artificially flat. With more long tasks in the mix, measured improvement actually accelerated.

But the benchmark's own architecture reveals the constraint. Of those 31 long tasks, only five have measured human baseline times; the rest use estimates. Building tasks that take expert humans eight-plus hours to complete, with verified completion criteria, is extraordinarily expensive. And benchmark tasks are simplified proxies for what models actually need to learn: sustained, multi-step reasoning across real professional workflows.

Meanwhile, Epoch AI projects that the stock of high-quality public text data will be fully utilized between 2026 and 2032. General-purpose web text is approaching exhaustion. What remains scarce is domain-specific, long-horizon human demonstration data: the recorded workflow of a nurse navigating a prior authorization denial.

What a five-hour task looks like in healthcare

Consider what a 320-minute time horizon means against the operational reality of a health system. The AMA's 2025 survey of 1,000 physicians found that practices complete 39 prior authorization requests per physician per week, consuming 13 hours of physician and staff time. Beginning January 2026, CMS requires payers to respond to standard prior authorization requests within seven calendar days and expedited requests within 72 hours, cutting existing timeframes roughly in half.

A single complex prior authorization (assembling clinical documentation, matching it to payer-specific criteria, submitting, receiving a denial, preparing an appeal) is precisely the kind of multi-hour workflow that falls within the current frontier model's time horizon. KFF's analysis of CMS data found that Medicare Advantage insurers made nearly 53 million prior authorization determinations in 2024, denied 4.1 million, and that 80.7% of appealed denials were overturned. Each overturned denial represents a workflow executed twice. That is the kind of sustained, document-intensive reasoning that models are approaching the ability to perform, but only if trained on demonstrations of clinicians actually doing it.

Banking runs on the same clock

The same clock runs through banking. The work that carries the most risk is rarely a single classification; it is the commercial credit review, the anti-money-laundering investigation, the complex fraud case that unfolds across hours or days. An analyst pulls transaction histories, reconciles counterparties, weighs the exceptions that do not fit the rule, and documents a decision chain that has to survive examination. The reasoning lives in the transitions, the judgment calls about what to escalate, what to hold, and what to clear. These are precisely the long-horizon, multi-step workflows the time-horizon curve says models are now approaching, and precisely the work no public dataset has ever captured. The demonstrations that would teach a model to do them sit inside banks, in the people who do them, not in any training corpus.

The annotation pipeline was not built for this

Most of the human feedback infrastructure that powers current model training was designed for a different era of AI capability. RLHF pipelines optimized for short preference pairs, choosing between two chatbot responses in under a minute, cannot produce the training signal that a model needs to learn eight-hour clinical documentation workflows. The mismatch is structural, not incremental.

A domain expert demonstrating a complex prior authorization workflow does not produce a data point in minutes; they produce it over hours of screen-captured work, decision narration, and exception handling. The cost per demonstration rises by orders of magnitude. And the expertise is not generic: it requires people who have actually worked prior authorization denials under production conditions.

That expertise is rarely inside the organization buying the model. Health systems and banks have operational staff who perform these workflows, but capturing them as structured training data, with verified completion criteria, is a different discipline. It requires egocentric data capture infrastructure, episode-level annotation rather than sentence-level annotation, and quality assurance that can validate multi-hour task completions against production standards.

The trendline bends where the data runs out

METR's researchers are explicit about the ceiling: even TH1.1's expanded suite has few tasks that the latest models cannot perform. The evaluation infrastructure is racing to keep up. But the same dynamic applies to training: models that sustain five hours of autonomous work need five-hour training episodes. Models approaching workday-length capability need workday-length demonstrations.

Labs without libraries of long-horizon human demonstrations (recorded, annotated, domain-verified workflows from healthcare and banking operations) will find their models flattening on the trendline. Not because the architecture has hit a wall, but because the training signal has. The compute exists. The algorithmic capacity exists. What does not exist, at scale, is the captured expertise of domain professionals performing the multi-hour tasks these models are now capable of learning.

The organizations that will matter in the next phase of this curve are those with the infrastructure to capture, annotate, and validate long-horizon expert demonstrations across regulated industries, at a quality level that sustains the 89-day doubling time rather than slowing it. That data infrastructure problem cannot be solved with general-purpose annotation platforms or crowdsourced preference labels. It requires domain depth, regulatory fluency, and the operational footprint to embed capture workflows inside production environments where prior authorization denials actually happen.