AI coding is transferring from a “software dividend” to a full system rebuild. For 2 years probably the most tangible change builders felt was quicker autocomplete and fewer boilerplate; that single-point effectivity part is over. Fashions now learn whole codebases, perceive mission construction, and be a part of pull requests and code evaluation — shifting software program improvement from “human-led, AI-assisted” towards “human units the aim, AI runs the method.” Cursor, the AI coding firm that has ridden the sector’s swings on its solution to a roughly $2 billion income run price, has now put exhausting numbers behind that shift in its first Developer Habits Report, drawn from its personal product information throughout 5 themes.
For any developer or engineering chief deciding learn how to use these instruments, the report’s worth is that it replaces vibe with proof — and the proof carries a warning most protection misses: AI is widening the hole between builders, not closing it.
How A lot Sooner Are Builders Truly Working?
By Cursor’s measure, the speed at which builders write code has roughly doubled yr over yr, and the tempo has stored accelerating because the begin of 2026. Strains of code are an imperfect metric, however they offer a directional learn — and the form of the work is altering, not simply the quantity. Strains added per pull request are up about 2.5 instances yr over yr, with progress nonetheless climbing, and “mega” pull requests that change at the very least 1,000 strains have gotten widespread as builders use AI to tackle bigger models of labor directly; Cursor notes a January 2026 bounce when many customers had been attempting the most recent fashions.
AI periods are additionally getting deeper. The common variety of software calls per AI dialog rose roughly 30% during the last two months, as assistants extra typically learn and edit recordsdata, search code, run command-line instructions, and browse the online. And the output is sticking: the share of accepted AI-suggested code nonetheless current 60 minutes later climbed from about 76% to 81% since early 2026 — an indication extra of what the mannequin writes genuinely survives evaluation quite than being reverted.
👉 Learn extra:
Cursor Composer 2.5 Matches Claude Opus 4.7 on Coding Benchmarks at One-Tenth Value
Why Does the Mannequin You Choose Value So Otherwise?
As fashions develop extra succesful and tackle deeper duties, value turns into a bigger a part of the product expertise, and the trade-offs are stark. Throughout seven mannequin households, the per-request value differs by practically 9 instances, that means the an identical workflow can value wildly totally different quantities relying on the mannequin behind it.
Uncooked worth, although, is deceptive. When Cursor measured the code that really survives — the strains accepted and stored — the biggest hole between low-cost and costly fashions narrowed to about 7 instances, as a result of pricier fashions have a tendency to provide extra usable code per request and so are usually not as costly as they first look. Plotting every mannequin’s rating on Cursor’s inside analysis, CursorBench, in opposition to its common activity value produces a transferring “cost-quality frontier,” the sensible map a group makes use of to determine which mannequin is price its worth for a given job.
Is AI Closing the Hole Between Builders, or Widening It?
Right here the report cuts exhausting in opposition to a typical assumption. Many anticipate AI to flatten the distinction between sturdy and weak builders; Cursor’s information exhibits the other, at the very least up to now — it amplifies the benefit of one of the best.
AI utilization is awfully concentrated. A small fraction of builders account for many AI-generated code, AI spending, and token consumption, with Gini coefficients of 0.77, 0.75, and 0.72 throughout these measures (the nearer to 1, the extra concentrated). The output hole is widening in absolute phrases: the highest 1% of builders produce 46 instances as many strains of code because the median lively person and 15 instances as many merged commits, whereas builders on the ninetieth percentile lead the median by a much smaller margin. The sample matches the instinct that builders who perceive structure, decompose duties properly, and may choose mannequin output high quality flip AI into leverage — whereas those that deal with it as a question-and-answer field see restricted beneficial properties. The sensible takeaway for a reader: the ability that compounds in an AI workflow is judgment, not typing pace.
What Is Driving AI Coding’s Prices Now?
As fashions tackle tougher duties, they learn much more earlier than they write, ingesting the codebase, the person’s intent, and the encircling workflow. The ratio of enter to output tokens is rising rapidly — the mannequin does extra “homework” per token of code it generates. Enter tokens now make up greater than 90% of non-cached token utilization, and their share of equivalent-price token value has climbed from roughly half to just about 70% yr so far.
This shift is, counterintuitively, good for value, as a result of enter tokens are far cheaper than output tokens and cached reads are cheaper nonetheless. And caching now dominates: cache-read tokens account for the massive majority of whole token exercise, that means AI coding more and more reuses beforehand processed context as a substitute of re-reading every thing from scratch. That rising context capability tracks with the rising retention price — extra context helps the mannequin write extra correct, longer-lived code. Additionally it is why the true aggressive moat in AI coding is migrating from uncooked mannequin cleverness towards context administration, cache effectivity, and price management. AI coding is beginning to look much less like a wiser code editor and extra like a brand new software-production infrastructure.
How Far Is AI Coding Shifting Towards Automation?
AI coding instruments started by dashing up particular person builders; Cursor’s information exhibits the subsequent part has began, with AI turning into infrastructure for automating the entire software-development lifecycle. Because the begin of 2026, AI adjustments auto-accepted into commits with out line-by-line human evaluation have grown greater than fivefold — a direct measure of rising belief within the mannequin to run components of the workflow itself.
The automation is spreading throughout workflow varieties, too. Adoption of Cursor’s automation options is climbing, with safety evaluation rising as a very sturdy use case, and newer information exhibits early demand for SDK-driven runs as builders look to show Cursor’s AI infrastructure right into a programmable platform they’ll customise on demand. It’s early, however the first automation patterns are seen — they usually level previous code era towards evaluation, testing, and deployment.
👉 Learn extra:
Salesforce Will Not Rent Extra Software program Engineers Subsequent 12 months As Claude Code Compresses Migrations
What the Report Alerts
Taken collectively, the info maps the place AI coding sits and the place it’s heading. The early contests had been about mannequin functionality and interplay expertise — who generated extra precisely and responded quicker. As duties develop extra advanced, the sturdy benefits are transferring towards context administration, caching, and price management, the unglamorous plumbing of a manufacturing system. For builders, the message is double-edged: AI is a robust lever, nevertheless it rewards those that already know learn how to wield it, and the hole between the highest and the center is widening. And as auto-accepted adjustments climb and automation spreads past code era, the human position is shifting from writing every line to setting targets and supervising a system that more and more runs itself.
Often Requested Questions
What’s the Cursor Developer Habits Report?
It’s Cursor’s first report on how builders use AI coding instruments, primarily based on the corporate’s personal product information and launched in spring 2026. It covers 5 themes: developer acceleration, the economics of mannequin alternative, the widening hole amongst prime customers, the rise of context and caching, and the transfer towards automation.
Does AI coding make all builders equally productive?
No. Cursor’s information exhibits the other: AI beneficial properties are extremely concentrated, with the highest 1% of builders producing 46 instances the code of the median person and 15 instances the merged commits. The report suggests AI amplifies the benefit of expert builders who can architect, decompose duties, and choose output high quality.
Why does AI coding value differ a lot between fashions?
Per-request prices range by practically 9 instances throughout the seven mannequin households Cursor analyzed. The hole narrows to about 7 instances when measured in opposition to code that really survives evaluation, as a result of costlier fashions typically produce extra usable code per request, offsetting a part of their greater worth.
What does “the rise of context” imply within the report?
Fashions now learn much more earlier than producing code: enter tokens exceed 90% of non-cached utilization, and cache reads make up most token exercise. This reuse of prior context lowers value and improves accuracy, and it’s shifting AI coding’s aggressive edge towards context administration, caching, and price management quite than uncooked mannequin intelligence.