Difficulty Frontier Task Generation

This bottleneck is really about generating work that still teaches a top model something new. Easy coding tasks are already saturated, because frontier models can patch small bugs and follow short issue threads, so the scarce input is harder tasks that require understanding a large repo, planning several steps, editing hundreds of lines, and surviving long rollout times during RL. Proximal is built around creating, grading, and training on exactly that kind of work.1 2 3 4 5

• Public coding benchmarks have skewed short. FrontierSWE argues SWE-Bench Pro solutions average only 107 lines of code, while its own tasks come from performance engineering, computational science, and ML research, where agents have to stay coherent across much longer chains of work. That shows why harder task generation, not just more labels, becomes the binding constraint.5
• The operational problem gets much tougher as task difficulty rises. Proximal describes RL steps on 10 hour tasks as roughly 60 times slower than RL steps on 10 minute tasks, which means every mistake in task design, grading, or sandbox reliability becomes expensive. The bottleneck is not only finding hard tasks, but making them trainable at scale.3 2
• This is a different business from broad data labeling. Scale, Surge, Mercor, and Micro1 all help supply expert data or evals, but Proximal positions itself around research driven task generation and post training infrastructure, with proof points like an 8.4 point DeepSWE Pass@1 improvement on its internal coding tasks and a reported $200M annualized revenue run rate by September 2026.6 7 2 4

The next leg of the market is moving from bug fixing copilots to agents that can own larger slices of software work. As that shift continues, the winners are likely to be the companies that can continuously manufacture new hard tasks, reliable verifiers, and long horizon training loops faster than model capability improves.1 3 4 5