The numbers are in, and they tell two different stories at the same time.
JetBrains’ April 2026 workplace survey found that 90 percent of developers now regularly use at least one AI tool at work. GitHub Copilot leads at 29 percent usage. ChatGPT follows at 28 percent. Claude Code and Cursor are tied at 18 percent each. That’s four tools splitting the developer market, with adoption rates that would have seemed absurd two years ago.
But Stack Overflow’s 2025 Developer Survey — the most comprehensive industry data available — found that while 84 percent of developers use or plan to use AI tools, only 29 percent trust their accuracy. That number hasn’t budged in the year since, despite tools getting measurably better. The gap between “I use this” and “I trust this” is the most important dynamic in AI-assisted development right now, and almost nobody is measuring it properly.
What 90 percent adoption actually means
A 90 percent adoption rate sounds like a done deal. But the number hides more than it reveals.
First, “regular use” is doing heavy lifting in that survey. JetBrains counted anyone who uses an AI tool at least once a week as a regular user. That’s a low bar. A developer who accepts one Copilot suggestion per day and ignores the rest counts the same as someone who builds entire features with AI assistance. The metric tells you the tools are present. It doesn’t tell you how much they matter.
Second, the 90 percent figure includes tools that many developers don’t think of as “AI coding tools.” ChatGPT at 28 percent usage likely includes developers who paste error messages into the chatbot for debugging help. That’s useful, but it’s a fundamentally different workflow than having an AI agent write production code. Lumping them together inflates the sense that AI has transformed how software gets built.
Third, awareness outpaces adoption. GitHub Copilot has 76 percent awareness among developers but 29 percent usage. That’s a 47-point gap. Developers know these tools exist. Many have tried them. A significant chunk decided they weren’t worth the friction. The question is whether that decision reflects the tools’ actual limitations or the developers’ resistance to changing established workflows.
There’s also a demographic split that the headline number doesn’t capture. Junior developers adopt AI tools at higher rates than seniors, partly because they have less invested in existing workflows and partly because AI suggestions fill gaps in their knowledge. Senior developers who’ve spent years building mental models of codebases are more likely to find AI suggestions disruptive rather than helpful — the tools suggest patterns that conflict with the senior developer’s established approach, creating friction instead of speed.
The result is that the 90 percent adoption number masks a real divide in how different experience levels interact with these tools. For juniors, AI is a teacher and accelerator. For seniors, it’s a sometimes-useful, sometimes-annoying colleague whose suggestions need constant evaluation. Both groups count as “regular users.”
The Claude Code anomaly
One number in the JetBrains data deserves its own section: Claude Code went from roughly 3 percent workplace usage to 18 percent in under a year. That’s a sixfold increase. No other tool in the survey grew that fast.
The obvious explanation is that Claude Code got better. Anthropic has shipped improvements steadily, and developers who tried it early and bounced off may have come back to find a different product. But there’s a less obvious factor: Claude Code’s growth tracks with the broader shift from autocomplete-style tools to agentic workflows.
GitHub Copilot started as autocomplete. It suggests the next line, the next function, the next block. Claude Code and Cursor position themselves differently — as tools that can understand a codebase, plan changes, and execute multi-file refactors with human oversight. The growth in Claude Code’s usage may reflect developers graduating from “AI helps me type” to “AI helps me think.”
If that’s the case, the 90 percent adoption number is misleading in a different way. It counts autocomplete users and agentic workflow users as the same category, when they’re using fundamentally different capabilities and getting fundamentally different value.
The distinction matters because the two categories have different failure modes. Autocomplete fails by suggesting wrong code that you catch in review. Agentic workflows fail by making architectural decisions that you don’t notice until the codebase is three refactors deep and something subtle breaks. The trust implications are different, the skill requirements are different, and the organizational impact is different. Collapsing them into a single metric makes the market look more uniform than it actually is.
Cursor, tied with Claude Code at 18 percent, takes yet another approach. It combines an AI-native code editor with model selection — developers can choose between different underlying models depending on the task. This flexibility appeals to developers who want control over which AI they’re working with, rather than being locked into a single provider’s approach. The Cursor-Claude Code tie at 18 percent suggests the market is fragmenting along workflow preferences rather than consolidating around a single winner.
The trust problem is a quality problem
The 29 percent trust figure from Stack Overflow hasn’t changed despite tools improving. Why?
Part of it is that expectations scale with capability. When Copilot suggested a single line of code, being wrong 70 percent of the time was annoying but manageable — you’d catch it in review. When Claude Code proposes a twenty-file refactor, being wrong means potentially introducing subtle bugs across an entire codebase. The stakes of trusting AI output increase as the tools take on larger tasks, so trust doesn’t increase even when accuracy does.
Part of it is the verification burden. AI tools shift work from writing code to reviewing code. That’s not necessarily faster. If you spend fifteen minutes writing a function yourself versus five minutes generating it with AI and ten minutes verifying it’s correct, you haven’t saved time. You’ve rearranged it. The trust problem persists because developers have learned, through experience, that verifying AI output takes real effort.
And part of it is the asymmetry of failure. When AI code works, nobody notices. When it breaks, it breaks in production on a Friday evening. Negative experiences stick longer than positive ones, and the 29 percent trust number reflects that accumulated memory.
There’s also the hallucination problem that nobody has solved. AI coding tools don’t just get things wrong — they get things wrong with confidence. A function that doesn’t exist, an API parameter that was deprecated two versions ago, a library import that doesn’t match the package version in your project. These aren’t rare edge cases. They’re routine. And each one erodes trust a little further, even when the tool gets the next ten suggestions right.
The companies building these tools know this. Copilot now includes inline confidence indicators. Cursor shows you the reasoning behind its suggestions. Claude Code can explain what it’s doing and why. But transparency about uncertainty is a double-edged sword: telling developers “I’m 60 percent sure about this suggestion” makes them less likely to trust it, even though it’s more honest than presenting every suggestion with equal confidence.
What the enterprise numbers reveal
Microsoft’s earnings data adds another layer. GitHub Copilot reached 140,000 organizations by Q3 FY2026, tripling year-over-year. Microsoft 365 Copilot surpassed 20 million paid seats, a 250 percent increase. Siemens deployed Copilot to 30,000 developers.
These are staggering adoption numbers at the enterprise level. But enterprise adoption and individual developer trust are different things. Companies buy Copilot licenses for the same reason they buy any productivity tool: the projected ROI justifies the cost, and not buying it puts you at a perceived disadvantage relative to competitors. Individual developers use it because it’s available and their manager expects them to.
The gap between “my company bought this for me” and “I trust this to produce correct code” is where the real friction lives. Enterprise adoption data measures purchasing decisions, not developer confidence. A company can deploy Copilot to 30,000 developers and still have most of those developers treating its output with skepticism.
What happens next
The AI coding tools market is heading toward a reckoning that the adoption numbers obscure.
The four-tool split (Copilot, ChatGPT, Claude Code, Cursor) will likely narrow. Tool preference in this category shifts fast — Claude Code’s sixfold growth in a year proves that. The tools that survive will be the ones that solve the trust problem, not just the ones that ship features faster.
Trust won’t come from better benchmarks or marketing claims. It’ll come from tools that explain their reasoning, flag their own uncertainty, and make verification easier rather than harder. The developers who use AI tools daily but don’t trust them aren’t luddites. They’re pragmatics who’ve learned that unchecked AI output creates technical debt faster than it creates value.
There’s also the question of measurement. The industry still doesn’t have a good way to measure whether AI tools actually make developers more productive. Self-reported surveys capture subjective impressions. Controlled studies like the one Microsoft ran with Copilot show modest gains on specific tasks but don’t capture the full workflow impact. Until someone figures out how to measure productivity across the entire software development lifecycle — not just the coding phase — the productivity claims from tool vendors will remain unverifiable.
The pricing picture is shifting too. GitHub Copilot costs ten dollars per month for individuals. Claude Code and Cursor charge more. As competition drives prices down and features up, the economic calculus changes. A tool that costs thirty dollars per month needs to save more than thirty dollars worth of developer time to justify itself. For senior developers whose time is expensive, that bar is lower. For junior developers, it’s higher. The trust gap may ultimately be resolved by economics as much as by technology.
The 90 percent adoption rate will keep climbing. The 29 percent trust number will be the one to watch. When those two numbers converge — when developers trust what they’re using at roughly the same rate they’re using it — that’s when AI-assisted development will have actually arrived. Until then, the industry is running on adoption without confidence, and that’s a gap that eventually closes one way or another.