OpenAI Is Slowing Hiring. Anthropic's Engineers Stopped Writing Code. Here's Why You Should Care.

AI capability surged in late 2025—so fast that many workplaces haven’t adjusted their workflows yet—creating a widening gap between what cutting-edge models can do and how most knowledge workers actually use them. OpenAI CEO Sam Altman admitted he still runs his own work much the same way, even as internal benchmarks and external reports suggest frontier systems now outperform human experts on a majority of well-scoped knowledge tasks. The practical result is “capability overhang”: teams with access to the same tools often operate at an older level of usage, while power users shift into long-running agent loops and task orchestration.

The turning point is framed as a December “phase transition,” not a single model release. Within about six days late last year, multiple frontier models landed—Google’s Gemini 3 Pro, OpenAI’s GPT 5.1 Codex Max (followed by GPT 5.2), and Anthropic’s Claude Opus 4.5. These releases share a theme: sustained autonomous work over hours or days rather than minutes. GPT’s 5.1/5.2 class models are positioned for continuous operation, Claude Opus 4.5 adds an “effort” control for reasoning intensity, and both ecosystems push techniques like context compaction so models can summarize their own work and maintain coherence over longer sessions.

But better models alone weren’t enough. The real unlock came from orchestration patterns that spread quickly in late December and early January. One was “Ralph,” a minimalist bash-loop approach created by Jeffrey Huntley that repeatedly runs Claude Code with git commits and file-based memory, wiping context windows when they fill and continuing until tests pass. Instead of elaborate multi-agent handoffs, persistence and looping proved more reliable—especially when models keep pausing, asking permission, or losing the thread.

A second viral pattern, “Gas Town,” coordinated dozens of agents in parallel through a workspace manager built by Steve Yagi. Despite its maximalist design, it reinforced the same core insight: the bottleneck shifted from “can the model write code?” to “can a human scope tasks and manage the right number of agents productively?” In late January, Anthropic’s Claude Code task system made that idea more native. Rather than forcing one long conversation to hold everything, the task system treats dependencies structurally: tasks can spawn isolated sub-agents with fresh context windows, and completion automatically unblocks dependent work. The result is a to-do-list-like interface that can coordinate 7–10 sub-agents at once, selecting different model sizes (e.g., Haiku for quick searches, Sonnet for implementation, Opus for deeper reasoning) while preventing cross-contamination of context.

This shift helps explain why hiring is slowing. Altman said OpenAI plans to dramatically reduce hiring pace because AI tooling expands the effective span of existing engineers; new hires are asked to do work that would normally take weeks in 10–20 minutes using AI tools. The transcript ties this to benchmark movement: GPT thinking tied or beat humans on 38% of well-scoped tasks earlier, rising to 74% with GPT 5.2 Pro.

Closing the overhang requires changing how work is managed. Power users move from asking questions to writing declarative specifications with success criteria, accept that agents will fail and iterate (as Ralph does by retrying until tests pass), and invest more in reviews, evals, and tests than in manual implementation. The biggest risk isn’t model weakness; it’s supervision and management—teams can generate large volumes of plausible but wrong work if they don’t scope tasks well and monitor outcomes. The forecast is that as orchestration becomes standard infrastructure and agent loops run longer, the ceiling lifts for complex software work, making “prompting” feel increasingly outdated compared with running parallel autonomous task systems.