14 Hours From a Single Prompt: The Steepest Trend in AI History

14 Hours From a Single Prompt: The Steepest Trend in AI History

There is a graph circulating in AI circles right now that is getting more attention than almost any other. It comes from METR — an independent nonprofit research institute in Berkeley that evaluates frontier AI models on their autonomous capabilities. And what it shows is striking: the ability of AI agents to autonomously complete long tasks has been growing exponentially — consistently, for six years.

What the METR Time Horizon Measures

METR defines the 50% time horizon as the task length (measured in human expert time) at which an AI agent succeeds in roughly half of attempts. Tasks are drawn from three domains: software engineering (HCAST, SWAA), ML research engineering (RE-Bench), and cybersecurity. They are deliberately realistic and messy — no memorization, no pattern matching. An agent gets an isolated environment, a goal, and has to figure out the rest.

For each model, METR fits a logistic curve that predicts success probability as a function of task length. The point where this curve intersects the 50% line is the time horizon. The R² value of the exponential trend fit across all models is 0.93 — this is not noise, it is a robust pattern.

The Concrete Progression: Model by Model

Here are the measured 50% time horizons for the key frontier models from METR's current TH1.1 benchmark suite:

The progression is striking: from ~1 hour for Claude 3.7 Sonnet in February 2025 to 14.5 hours for Claude Opus 4.6 in February 2026 — a tenfold increase in twelve months.

METR's trend line shows a doubling time of approximately 123 days (~4 months) based on data from 2023 onward — significantly faster than the 7-month doubling time measured over the full 2019–2025 period.

Why Is the Trend Accelerating?

Three independent scaling axes are running in parallel and reinforcing each other:

1. Better base models. Each new model generation brings substantial improvements in long-horizon reasoning, multi-step planning, and error-resilient execution. Claude Opus 4.6 and GPT-5-class models demonstrate significantly more robust behavior on tasks requiring dozens of sequential decisions.

2. Inference-time scaling (extended thinking). Claude 3.7 Sonnet was one of the first models with genuine extended thinking. The ability to allocate more compute at runtime to difficult subproblems — without retraining the base model — caused a step change in time horizons. This explains the inflection in the trend curve in early 2025.

3. Better scaffolding tools. Claude Code, OpenAI Codex CLI, and similar agent frameworks have fundamentally improved how models interact with their environment: persistent bash sessions, structured tool use, error retry, file context management. METR uses these scaffolds directly in their evaluations (ReAct, Triframe, Claude Code). Scaffolding quality directly influences measured capabilities.

4. Competitive pressure. Anthropic, OpenAI, and Google shipped multiple frontier models between August 2025 and February 2026, sometimes within days of each other. The arms race is compressing release cycles and accelerating the curve.

What the Benchmark Measures — and What It Doesn't

An important caveat: METR tests cleanly scoped tasks with automated scoring. Agents have no prior context on the project and receive exactly the same resources as human baseline participants. This is methodologically cleaner than most benchmarks, but also more constrained than real-world work.

METR itself notes in their paper that performance drops substantially when tasks are evaluated holistically rather than algorithmically. In a follow-up study, they found that roughly half of the test-passing PRs that current agents produce on SWE-Bench Verified would not actually be merged by real maintainers. The gap between "passes automated tests" and "is good software" remains real.

Time horizons also vary substantially by task type. Agentic computer use lags software engineering tasks by roughly 50x according to METR's own cross-domain analysis. The 14.5 hours applies to the specific mix of SE, ML, and security tasks in the TH1.1 suite — not to arbitrary knowledge work.

What This Means for Software Developers

The tasks in the METR benchmark — complex debugging, ML experiments, security analysis — are precisely what senior engineers work on daily. And the models METR tests are the same ones deployed via Claude Code, Codex, and Cursor in production workflows today.

Claude Code reportedly saw 5.5x revenue growth by mid-2025, according to Anthropic. METR notes explicitly in their analysis that these tools are already building real engineering infrastructure — not as proof-of-concept, but as productive work.

The practical implication: the bottleneck is shifting. AI can increasingly take on implementation work. What remains — and gains in value — is what AI structurally cannot do: define system boundaries, evaluate trade-offs, build context, take responsibility.

Where Does the Curve Lead?

If the current trend (doubling every 4 months) holds, the extrapolations are:

• Mid-2026: ~60–80 hours (one working week)
• End of 2026: ~several weeks
• 2027–2028: Month-long autonomous projects

METR's 95% confidence intervals for these extrapolations are wide, and the trend could flatten. But there is currently no empirical sign of deceleration — if anything, recent data points are tracking above the trend line.

Worth noting is a feedback loop METR's researchers explicitly flag: improved AI agents are increasingly useful for developing even more capable AI agents. The tools METR tests are the same ones Anthropic and OpenAI use in their own research. This could produce superexponential growth — agents building agents building better agents.

Conclusion

The METR time horizon graph is remarkable precisely because it measures AI capability in a single, intuitively meaningful unit: human working hours. No abstract points system, no comparison with exam scores — just the plain question: how much autonomous, real engineering work can an AI agent do today?

The answer in March 2026: 14.5 hours. With a doubling time of four months.

Sources

• METR — Task-Completion Time Horizons of Frontier AI Models (live, continuously updated): metr.org/time-horizons
• METR — Measuring AI Ability to Complete Long Tasks (original paper, March 2025): metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks
• METR — Time Horizon 1.1 (updated benchmark suite, January 2026): metr.org/blog/2026-1-29-time-horizon-1-1
• Kwa et al. (2025) — arXiv paper: arxiv.org/abs/2503.14499
• METR — Many SWE-bench-Passing PRs Would Not Be Merged (March 2026): metr.org
• The AI Digest — A new Moore's Law for AI agents: theaidigest.org/time-horizons