In one year, Claude Code went from a beta without plan mode to a million-token context and 9 subagents

The market for AI development tools changed faster in a year than many expected. What looked like a raw beta in March 2025 had turned, by spring 2026, into agentic systems with million-token context, subagents, and sharply cheaper economics of use.

From beta to platform

In March 2025, Claude Code was a beta and, according to the author of the review, could not even work in plan mode. This matters not as a nitpick about one feature but as a marker of maturity for the whole category. Back then, AI coding was still seen as a helper for individual fragments of code: suggest something, finish a line, fix a test. A workflow in which the tool first builds a plan, then distributes work among several workers and holds a large context, was not yet the norm.

Over the following year the picture changed radically. Claude Code gained a context of one million tokens and nine subagents, and alongside it grew a whole class of CLI tools that simply did not exist at the start of the period. Codex CLI, for one, did not yet exist. It shows how quickly the developer's AI environment moved from point-by-point autocomplete to semi-autonomous engineering work, where the model no longer just generates code but also helps organize the process.

The numbers for the year

The most telling changes show up in the numbers. The author compiled a year of data and found a picture in which not only model capabilities are growing, but also the scale of the business around them. If at the start of the period the top model solved about 65% of tasks on SWE-bench, the market then began moving along several axes at once: quality, context, cost, and agentic capability. It is the combination of these factors that makes the shift systemic rather than a local update to a few products.

• Claude Code went from a beta without plan mode to a mode with a million-token context
• The number of subagents grew to 9, which changes the very approach to decomposing tasks
• Context windows in leading tools swelled roughly 5x in a year
• The price of frontier coding fell roughly 16x, lowering the entry barrier for teams
• Cursor, according to the review, grew to $2 billion in revenue

Together, these metrics show the market developing in two directions at once. On one hand, models are getting stronger and better at sustaining a long engineering chain. On the other, access to these capabilities is getting cheaper, so agentic workflows reach not only large labs but ordinary product teams faster. A year ago, this combination was available mostly to early adopters and research groups.

The agentic leap

The strongest signal is not even the million tokens or the falling model prices, but the shift to agentic work at scale. Cursor, if the cited data is to be believed, is already running a thousand agents in parallel. This changes not only development speed but the structure of the work itself: some tasks are now solved not by a single chat with a model but by a swarm of specialized workers that research, write, test, and verify results in parallel.

One case stands out: a system of this kind wrote a browser in a week without human involvement. Even if you treat it as a demonstration of an extreme scenario rather than a standard process for any team, the conclusion is still stark: AI development is ceasing to be a tool for speeding up individual engineers and is starting to compete with the conventional organization of entire product cycles. Against this backdrop, the human role shifts toward setting tasks, checking quality, and choosing architectural boundaries.

What this means

Between March 2025 and spring 2026, AI coding moved from a "useful assistant" mode to an "operating system for development" mode. For teams, this means two things: the cost of experimentation falls, while the demands on decision-making speed rise. If the trend holds, the winners will not be those who merely have access to a model, but those who learn faster to manage agents, context, and verification of results.