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

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

From Beta to Platform

In March 2025, Claude Code existed in beta status and, according to the review author, could not even work in plan mode. This matters not as a quibble over a single feature, but as a marker of maturity for the entire category. Back then, AI coding was still perceived as an assistant for individual code snippets: suggest, complete, fix a test.

The scenario in which a tool first builds a plan, then distributes work among multiple executors and maintains large context, was not yet the norm. Over the next year, the picture changed radically. Claude Code gained million-token context and nine subagents, and a whole class of CLI tools grew alongside it, which simply did not exist at the beginning of the period — the same Codex CLI did not exist then.

This shows how quickly the AI environment for developers has shifted from point-based auto-completion to semi-autonomous engineering work, where the model not only generates code but also helps organize the process.

Figures Over a Year

The most telling changes are visible in the numbers. The author collected data over a year and got a picture where not only the capabilities of the models are growing, but also the scale of business around them. If at the beginning of the period the top model covered about 65% of tasks on SWE-bench, then the market began to move along several axes at once: quality, context, cost, and agency. It is precisely this combination of factors that makes the shift systemic, rather than a local update of a few products.

• Claude Code went from beta without plan mode to a mode with million-token context
• The number of subagents grew to 9, changing the very approach to task decomposition
• Context windows of leading tools expanded roughly 5-fold over the year
• The price of frontier coding fell roughly 16-fold, lowering the entry barrier for teams
• Cursor, according to the review data, grew to $2 billion in revenue

Together, these metrics show that the market is developing in two directions at once. On the one hand, models become stronger and better maintain long engineering chains. On the other, access to such capabilities becomes cheaper, which means agent scenarios reach not just large labs but regular product teams faster. A year ago, such a combination was available mainly to early adopters and research groups.

The Agent Leap

The strongest signal is not even a million tokens and not the cheapening of models, but the transition to mass agent work. Cursor, if the provided data is to be believed, is already running a thousand agents in parallel. This changes not only the speed of development, but also 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 executors that in parallel investigate, write, test, and verify the result.

What stands out separately is a case where such a system wrote a browser in a week without human involvement. Even if you perceive it as a demonstration of a limiting scenario rather than a standard process for any team, the conclusion is still stark: AI development is ceasing to be a tool for accelerating individual engineers and is beginning to compete with the familiar organization of entire product cycles. Against this backdrop, the human role shifts toward task setting, quality verification, and the choice of architectural frameworks.

What This Means

Over the period from March 2025 to spring 2026, AI coding transitioned from "useful assistant" mode to "operating system for development" mode. For teams, this means two things: the cost of experiments falls, but the demands on decision-making speed rise. If the trend continues, the winners will not be those who simply have access to the model, but those who quickly learn to manage agents, context, and result verification.