Claude Code Adoption Doubled. Copilot and Cursor Both Lost Ground.

Claude Code adoption jumped from 18% to 39% in seven months per JetBrains' 2026 survey, while Copilot and Cursor both lost ground.

Claude Code adoption more than doubled in seven months, from 18 percent of professional developers worldwide in January 2026 to 39 percent by May to July 2026, according to JetBrains' Developer Ecosystem Survey 2026, and it did not happen in a vacuum: GitHub Copilot's adoption fell 8 points and Cursor's fell 6 points over the same stretch. In the United States specifically, Claude Code adoption reached 47 percent, and 31 percent of developers now name it the tool they use most for coding, ahead of every other option in the survey.

Sources: JetBrains: AI Coding Agents, Adoption Trends (Developer Ecosystem Survey 2026)

The survey polled more than 15,000 professional developers worldwide, fielded from May to July 2026 and published in August 2026. The May-to-July figures come from JetBrains' Developer Ecosystem Survey 2026, its 10th annual edition; the January 2026 baseline comes from a separate JetBrains instrument, the AI Pulse survey, which JetBrains weights against the Ecosystem Survey's own population so the two are comparable, not one continuous panel run twice. In that one window, every major coding tool tracked in the survey moved, and not all of them moved up.

Sources: JetBrains: AI Coding Agents, Adoption Trends (Developer Ecosystem Survey 2026)

A disclosure before the numbers get more specific: we build TLM Forge, a process layer for AI coding designed to work the same way regardless of which tool or model sits underneath it. An argument against betting a team's process on one vendor's specific tool points somewhere commercial, so read the survey data as the load-bearing part of this post, and read the product section near the end knowing where it comes from.

What moved between January and July

All four major coding tools tracked in this survey moved in the same seven-month window. Here is the full scoreboard, January 2026 against May to July 2026:

Sources: JetBrains: AI Coding Agents, Adoption Trends (Developer Ecosystem Survey 2026)

Two things stand out beyond the top-line swap. First, Claude Code's rise is not a single-survey fluke: JetBrains' own tracking already had it near 3 percent of developers in the April-to-June 2025 window, rising to 18 percent by January 2026, before this survey caught it at 39 percent in May to July 2026. That is three consecutive waves of acceleration, not one favorable quarter. Second, adoption and preference are different measurements, and the gap between them matters. Being used by 39 percent of developers is not the same as being the tool 31 percent of developers reach for first, ahead of every other option. Claude Code leads on both counts here, but the distance between "people who use it" and "people who use it most" is exactly the kind of detail a single headline percentage erases.

Sources: JetBrains: Which AI Coding Tools Do Developers Actually Use at Work?, JetBrains: AI Coding Agents, Adoption Trends (Developer Ecosystem Survey 2026)

The same survey found that 90 percent of professional developers now use an AI coding agent at work at least weekly, and 68 percent use one daily. Put those two facts next to the tool-level numbers and a detail falls out on its own: Claude Code is used by 39 percent of developers but named primary by only 31 percent, which only works arithmetically if a meaningful share of developers are running it alongside something else rather than treating it as an exclusive choice. Multi-tool use is not a hypothetical here. It is close to the default.

Sources: JetBrains: AI Coding Agents, Adoption Trends (Developer Ecosystem Survey 2026)

Seven months is not a moat

GitHub Copilot led adoption among these four tools in January 2026, at 29 percent, ahead of Claude Code and Cursor, which were tied at 18 percent. By May to July 2026 it had fallen to second place behind Claude Code. Cursor's fall was proportionally sharper: tied for second in January, it finished last of the four by summer, even though its point loss was smaller than Copilot's. Rank order and point change tell different stories here, and both moved against the two tools that led seven months earlier.

Sources: JetBrains: Which AI Coding Tools Do Developers Actually Use at Work?, JetBrains: AI Coding Agents, Adoption Trends (Developer Ecosystem Survey 2026)

None of this argues for freezing a team's tool choice or ignoring whichever product is winning this quarter. It argues for separating the parts of a workflow that do not care which vendor sits underneath them from the parts that do. Some investments age badly when the leaderboard reshuffles every seven months:

• Prompt libraries and slash commands hand-tuned to one tool's specific syntax, which do not transfer when the team moves to a different agent.
• A training budget spent certifying engineers on one vendor's interface, rather than on how to write a spec, review a diff, or judge test coverage, skills that hold regardless of which tool is typing the code.
• CI and pre-commit hooks wired to one CLI's exact flags and output format, which break quietly the day that interface changes or a second tool joins the pipeline.
• Onboarding docs and internal wikis built around one tool's screenshots and menus, which go stale on roughly the same cycle this survey shows the market moving.
• Multi-year seat licenses or procurement contracts signed before the next wave of adoption data has a chance to land.

The alternative is not less investment, it is investment aimed one layer up. A defined spec review before code starts, a clear standard for what "done" means on a change, a rule that nothing merges without an independent reviewer, and a test suite that exists before the implementation does: all of that holds regardless of whether the agent behind it is this quarter's leader or next quarter's. Put the budget there, and a vendor swap costs a reconfiguration, not a rebuilt playbook.

A team starting from zero should still weight the current leader heavily: 39 percent adoption and a first-place primary-tool ranking are real signals, not noise. Treat that leader as the current best default, not a permanent architectural decision, so next year's leader slots into the same review and testing discipline without a rebuild.

The Codex gap: aware of it, not using it

OpenAI Codex's awareness climbed from 27 percent of developers in January 2026 to 65 percent by May to July 2026, while its adoption grew from 3 percent to 16 percent over the same window. Put plainly, about two in three developers now know what Codex is, but fewer than one in six have actually used it. Adoption did grow, more than five times over, so this is not a tool developers are ignoring outright. It is a tool most of the people who have heard of it have not yet tried.

Sources: JetBrains: AI Coding Agents, Adoption Trends (Developer Ecosystem Survey 2026)

That pattern is not unique to Codex. GitHub Copilot's own awareness sits at 79 percent against 21 percent adoption in the same survey, a comparable gap running the other direction: a tool nearly every developer has heard of that a minority actively uses. Do the arithmetic on both and the conversion rate lands in the same neighborhood: roughly one in four developers who know about either tool has actually adopted it. JetBrains' survey does not report why, and this post will not guess at a specific cause for either gap. What the data does support is naming the pattern plainly: a large share of developers knowing a tool exists is not the same as a large share choosing to run it, for the newest entrant and the long-established incumbent alike.

Sources: JetBrains: AI Coding Agents, Adoption Trends (Developer Ecosystem Survey 2026)

Build the layer that does not need to win the survey

The parts of a workflow least exposed to this kind of reshuffle are the ones defined above the tool: what a change is supposed to do before anyone writes code, who reviews it and how independently, whether tests exist before the implementation does, and what has to be true before a diff is allowed to ship. TLM Forge is built at exactly that layer. A spec audited before code starts, an independent multi-agent review of the plan and the diff, phase-gated TDD, and a convergence gate that blocks shipping until critical issues hit zero: none of that logic changes depending on whether the agent typing the code is Claude Code, Copilot, Cursor, or Codex. The discipline is portable in a way no single tool's specific quirks are, which is a practical answer to a market that has already reshuffled its leader three survey waves running.

Session memory has the same portability problem, at a smaller scale. A team that switches its primary coding tool, or simply runs two or three in parallel the way a growing share of developers already do, loses whatever project context lived inside the old tool's chat history. That context has to live in the codebase and the team's own documentation, not in any single tool's memory, since no vendor is guaranteed to still be the primary tool a year from now. MemX, from the same team behind TLM Forge, is a private, persistent memory layer for a different, personal problem: an individual's own photos, documents, voice notes, and messages, not project architecture.

Two related questions tend to come up once a team takes this seriously. How to choose a model for AI coding covers the trade-offs for picking a model per task rather than defaulting to whichever tool won the last survey, and seven common mistakes teams make with AI coding tools covers the process failures that show up regardless of which tool is in the seat, from merging diffs nobody read to losing project context between sessions.

Seven months from now another survey will show another set of numbers, and there is no reliable way to know today which tool will have gained the most or lost the most. What this round of data supports is narrower and more useful: the distance between today's leader and today's last place is not as wide, or as durable, as it looks from inside a single quarter. Plan the parts of the workflow that outlast the leaderboard, and let the tool underneath it change as often as the data says it will.