Developers are adding more code per week, with growth accelerating since the start of 2026. While this is not a perfect metric, it provides a directionally interesting baseline for understanding how developer work is changing.

Lines added per PR are up roughly 2.5x year over year and the growth rate is accelerating.

Mega PRs, defined as PRs with at least 1,000 lines changed, are becoming more common as developers use AI to take on larger units of work in a single PR. It's interesting to see the jump in mega PRs in January 2026, when many developers were trying out the latest improvements in coding agents and models.

In just the last two months, average tool calls per session have risen roughly 30%. Coding agents are taking on more complex work, reading and editing files, searching code, running shell commands, and browsing the web more frequently.

Since the start of 2026, the share of accepted AI lines still present after 60 minutes has risen from roughly 76% to 81%.

Cost per agent request varies by nearly 9x across model families, showing that the same workflow can have very different cost profiles depending on the model behind it.

Cost per accepted line varies by roughly 7x across model families, compared with nearly 9x for cost per request, suggesting that higher-cost models partially make up the difference by producing more accepted code per request.

This benchmark view plots model performance on Cursor’s internal eval suite, CursorBench, against average task cost, showing where models sit on the cost-quality frontier.

AI usage is highly concentrated, with a small share of developers accounting for a large share of AI lines, spend, and token consumption.

The Lorenz curves show this concentration, with Gini scores of 0.77, 0.75, and 0.72 across the three metrics, where higher scores on a 0-to-1 scale mean activity is more concentrated among fewer users.

We see p90 developers pulling farther away from median developers in absolute lines added per week, with p99 users even farther out in the tail.

Here’s another view of how dramatically the power-user gap widens at the tail.

P99 developers produce 46x more lines than the median active user and merge 15x more PRs than the median active PR author, while p90 users show large but much smaller gaps.

The ratio of input to output tokens is rising quickly, showing that models are consuming much more context for every token they produce. This shift suggests that models are doing more work up front before generating code.

The same relative shift toward input tokens shows up in the token mix. Input now accounts for more than 90% of input-output token volume, making context the dominant part of non-cache model usage.

Input tokens dominate token consumption, but their effect on cost is moderated by their lower unit price.

Even so, input tokens have become the majority of price-equivalent token costs, rising since the start of the year from roughly half of input/output token costs to nearly 70%.

The context story becomes even larger once cache is included. Cache-read tokens dominate total token activity, showing how much agent work now depends on reusing prior context rather than reading everything from scratch. We continually improve our agent harness to best cache tokens across models and providers.

Since the start of the year, more than 5x as many agent-generated changes are reaching commits without a separate manual diff acceptance step, suggesting that developers are trusting agents to carry more work through the commit flow.

It is still early, but the first autonomy patterns are coming into focus. Adoption of Cursor Automations is growing quickly, with security review emerging as a strong automation use case.

Even more recently, SDK runs show early demand for turning Cursor’s agent infrastructure into a programmable platform customized to how each company builds software.