AI Coding: what is getting real adoption now?
In our AI code assistant market deck, you will find everything you need to understand the market
SUMMARY
AI Coding: what is getting real adoption now? Agentic coding with human supervision has clearly crossed into real adoption, with developers using agents daily, companies paying for them at scale, and AI taking over meaningful parts of implementation, testing, review and maintenance.
The useful adoption question has shifted from whether developers use AI at all to how much work they hand over. Basic AI usage is already so widespread that frequency, delegation depth and enterprise spending now tell us far more than simple user counts.
Coding agents have moved unusually fast. JetBrains found 90% of professional developers using them at least weekly and 68% using them daily, which puts agentic workflows much closer to normal software development than to an experimental side tool.
The market is also reshuffling quickly. Claude Code currently leads professional developer preference, Codex has the sharpest recent growth curve, and GitHub Copilot still owns a huge distribution advantage through GitHub and enterprise engineering environments.
Claude Code’s position looks stronger than a survey fad because developer preference, repeat usage, revenue and enterprise spending are moving in the same direction. Enterprise customers now account for more than half of Claude Code revenue, which is a much harder adoption test than individual developer enthusiasm.
Codex remains smaller in professional share but looks unusually deep among its core users. Its adoption rose more than fivefold in roughly half a year, and a high share of primary Codex users say agents generate most of their code.
Developers are handing over far more than autocomplete. JetBrains respondents attributed about 47% of their recent code to full agent generation on average, while roughly one in five said agents generated more than 80% of their code.
The productivity story is less clean than the adoption story. Developers clearly keep using AI because it saves effort in enough situations, but controlled experiments and engineering telemetry still show that faster code generation can create slower review, more bugs, more incidents and more rework downstream.
The most successful use cases currently have tight feedback loops: implementation, tests, refactoring, migrations, debugging and review. Deployment, monitoring, architecture and ambiguous planning remain much harder to delegate because mistakes are costlier and correctness is less obvious.
The bigger shift is that software creation is expanding beyond traditional developers at the same time. Tools such as Replit and other AI app builders are turning prompts into working applications for founders, designers and product people, while professional engineers increasingly spend more time steering, checking and deciding what ships.
This market map, featured in our AI code assistant market deck, highlights top companies and startups in the AI code assistant market
What counts as real AI coding adoption now?
Real AI coding adoption today means developers keep using the tool for actual work, companies keep paying for it, and AI is trusted with meaningful chunks of the software process.
Basic usage numbers have become almost useless on their own. Google’s 2025 DORA study found that 90% of nearly 5,000 technology professionals were using AI at work. JetBrains later found that 90% of professional developers used at least one AI tool regularly at work. At that level, asking whether developers have “tried AI coding” tells us very little.
The more useful evidence is behavioral. Are developers using agents every day? Which tool becomes their main one? How much code are they actually handing over? Are companies buying enterprise access after pilots? Does AI move beyond autocomplete into tests, refactoring, review and larger tasks?
With that stricter bar, the picture is clearer. Autocomplete has already settled into normal development. Coding agents are rapidly joining it. Giving AI responsibility for an entire production system is still much rarer.
Are coding agents actually mainstream among developers now?
Coding agents have become mainstream among professional developers surprisingly fast, with daily use already common.
JetBrains surveyed more than 15,000 professional developers between May and July 2026 and found that 90% used coding agents at work at least weekly. Some 68% used them every day. The survey was reweighted across geography, employment status, programming language and JetBrains familiarity, so this is one of the better broad measurements we have.
Compare that with Stack Overflow’s 2025 survey. At the time, 52% of developers either stayed with simpler AI tools or did not use agents, while 38% said they had no plans to adopt them. The definitions differ, so the movement is not a clean year-on-year comparison. Still, the change is hard to miss.
The product itself has changed alongside the behavior. Developers increasingly give an agent a task, let it inspect multiple files, edit code, run commands, read failures and try again. That is a much bigger delegation of work than accepting an autocomplete suggestion.
A 90% weekly figure probably includes plenty of light users. The 68% daily number is harder to brush aside. AI agents are now part of normal development for a large share of professional programmers.
As this chart shows, and as featured in our AI code assistant market deck, search interest in AI code assistants has increased significantly
Which AI coding tools are developers actually choosing?
Claude Code currently leads professional developer usage, while Codex is growing the fastest and GitHub Copilot still has enormous distribution.
JetBrains gives us a useful repeated view of the market. In January 2026, GitHub Copilot was used at work by 29% of professional developers, while Claude Code and Cursor were both at 18% and Codex sat around 3%.
By May–July, Claude Code had reached 39%. Codex climbed to 16%. Copilot stood at 21%, while Cursor dropped to 12%. Claude Code also became the main AI coding tool for 31% of all professional developers surveyed.
That reshuffling happened in roughly half a year. Developer preference is moving much faster than the installed base of traditional development tools normally does.
Survey share only tells part of the story, though. Microsoft now reports 50 million GitHub Copilot users, while OpenAI has reported more than four million weekly Codex developers. Different companies define users differently, so those totals cannot be ranked directly against JetBrains adoption percentages.
The market now has three serious poles: Claude Code for strong developer preference, Codex for very fast recent growth, and Copilot for massive GitHub and enterprise distribution.
| AI coding tool | Professional adoption in January 2026 | Professional adoption in May–July 2026 |
|---|---|---|
| Claude Code | 18% | 39% |
| GitHub Copilot | 29% | 21% |
| OpenAI Codex | ~3% | 16% |
| Cursor | 18% | 12% |
| Google Antigravity | ~6% | ~6% |
Why is Claude Code winning so many developers?
Claude Code is currently the clearest developer favorite because it made agentic coding useful enough to become somebody’s main tool rather than another assistant sitting beside the editor.
The speed of adoption is unusual. JetBrains measured Claude Code at roughly 3% professional adoption during spring 2025, 18% in January 2026 and 39% by its latest survey period. About four out of five professional users who had adopted Claude Code said it was also their most-used AI coding tool.
Commercial usage reinforces that survey result. Anthropic says Claude Code now exceeds a $2.5 billion annualized revenue run rate, more than twice its level at the beginning of 2026. Weekly active users have doubled over the same period.
More importantly, enterprise customers now generate over half of Claude Code revenue, while business subscriptions have quadrupled. Individual developers can make a new coding product fashionable very quickly. Getting companies to expand usage after security, procurement and budget checks is considerably harder.
Anthropic also cited an external analysis estimating that Claude Code authored around 4% of public GitHub commits worldwide, twice the estimated share one month earlier. Public GitHub is only one slice of software development, but a tool touching roughly one in 25 public commits has clearly escaped the early-adopter bubble.
Claude Code has reached the point where adoption shows up in four places at once: developer surveys, repeated usage, revenue and enterprise spending.
If you want more recent data on this point, please see our latest AI code assistant market report.
This chart, featured in our AI code assistant market deck, illustrates yearly VC funding for AI code assistant startups
Can OpenAI Codex catch Claude Code?
Codex can realistically challenge Claude Code because its current adoption curve is much steeper than its smaller professional share suggests.
JetBrains measured Codex at roughly 3% professional usage in January and 16% by May–July. That is more than a fivefold increase in around half a year. Awareness jumped too, rising from 27% to 65% among professional developers.
OpenAI’s own numbers moved at similar speed. The company reported more than three million weekly Codex developers in early April 2026. Two weeks later, the figure had passed four million.
There is another interesting detail inside the JetBrains data. Among developers who use Codex as their main AI coding tool, 42% said agents generate more than 80% of their code. For Claude Code users, the equivalent figure was 32%. Codex currently has fewer professional users, but its user base appears unusually comfortable with deep delegation.
OpenAI is also pushing Codex into companies through Codex Labs and partnerships with Accenture, Capgemini, Cognizant, Infosys, PwC and TCS. Cisco uses Codex across connected repositories, Ramp for code review, Virgin Atlantic for test coverage and Rakuten for incident response.
Claude Code still has the larger professional footprint. Codex now has enough growth, usage intensity and enterprise distribution to make that lead vulnerable.
Has GitHub Copilot really lost the AI coding race?
GitHub Copilot has lost its position as the clear developer favorite, yet it remains one of the strongest AI coding businesses in the world.
The developer-preference data has clearly moved against Copilot. JetBrains measured professional adoption falling from 29% in January to 21% in May–July while Claude Code and Codex grew quickly.
Microsoft’s latest numbers make it impossible to describe Copilot as a product in retreat. GitHub Copilot now has 50 million users. Copilot revenue accelerated by more than 60% quarter over quarter, and Microsoft says Business and Enterprise seats are still growing.
GitHub itself has become a major distribution channel for agents. Microsoft says one in three pull requests on GitHub now involves an agent. GitHub has 225 million users, and more than 90% of the Fortune 500 use the platform for AI-powered development.
So there are really two races here. Claude Code currently has stronger pull among professional developers choosing their preferred coding agent. Copilot owns a much larger installed distribution surface inside GitHub and corporate engineering environments.
A product can lose mindshare among early adopters and still become much larger commercially. Copilot appears to be doing exactly that.
If you want more recent data on this point, please see our latest AI code assistant market report.
This chart, featured in our AI code assistant market deck, breaks down Anyshpere’s playbook in AI code assistants
Is Cursor actually losing developers?
Cursor is losing relative usage share today, but a business already generating around $1 billion in annualized revenue still has very real adoption.
JetBrains measured Cursor falling from 18% professional work usage in January to 12% by May–July. China was particularly weak, dropping from 28% to 16%. At the same time, Cursor’s awareness among developers actually increased.
The likely explanation sits in how AI coding itself is changing. Cursor became popular by rebuilding the editor around AI. Claude Code and Codex made the coding agent more portable. A developer can increasingly use a powerful agent through a terminal, IDE integration or cloud environment without moving their whole workflow into a dedicated AI editor.
Cursor still has millions of users and has crossed roughly $1 billion in annualized revenue. Those economics are far beyond experimental software.
The pressure is real, though. As agents become the product developers care about most, owning the editor becomes less valuable unless the editor provides a meaningfully better place to run those agents.
Are companies really rolling out AI coding agents?
Enterprise adoption of AI coding agents is now substantial, and companies are moving beyond small developer pilots into recurring engineering workflows.
Claude Code offers one of the cleanest commercial measurements: enterprise use generates more than half its revenue, while business subscriptions have quadrupled since the beginning of 2026.
OpenAI says Codex has become one of its fastest-growing enterprise products. The company has documented customers using it across different pieces of engineering work: Cisco for reasoning across large repositories, Ramp for reviewing code, Virgin Atlantic for test coverage and Rakuten for incident response. OpenAI has also partnered with major systems integrators specifically to help larger companies deploy Codex.
GitHub brings much broader distribution. Microsoft now says over 90% of the Fortune 500 use GitHub for AI-powered development, with both Business and Enterprise Copilot seats continuing to grow.
Then there is Google, which gives us a glimpse of how deep adoption can become inside one unusually advanced engineering organization. Google says AI currently generates 75% of its new code before engineers review and approve it. The same company recently reported that an agent-assisted complex code migration finished six times faster than a comparable engineer-only process a year earlier.
Google is an extreme case, so 75% AI-generated code should not be treated as the normal enterprise benchmark. Taken together with Claude Code’s enterprise revenue, Copilot’s corporate reach and Codex deployments, the broader conclusion is already solid: companies are buying agents for actual software work at scale.
| Company or product | Recent adoption evidence | What developers are using it for |
|---|---|---|
| Claude Code | Enterprise generates more than half of revenue | Coding and agentic development |
| GitHub Copilot | 50 million users; Business and Enterprise seats still growing | Coding inside GitHub and IDE workflows |
| OpenAI Codex | More than 4 million weekly developers | Coding, review, testing, repository work and incident response |
| Google internal AI | 75% of new code AI-generated before engineer approval | Coding, migrations and multi-agent workflows |
This chart, featured in our AI code assistant market deck, illustrates yearly funding for AI code assistant startups
How much code are developers actually handing to AI?
AI now writes a large share of professional developers’ code, although complete delegation is still concentrated among a smaller group of heavy users.
JetBrains asked more than 15,000 professional developers to split the code they had produced during the previous month into fully agent-generated, AI-assisted and manually written code. On average, developers attributed roughly 47% to full agent generation.
More than half said manual-only coding represented less than 20% of their output. Around one in five said they wrote no code without some AI involvement.
The extreme end remains smaller. About 22% said agents generated more than 80% of their code. JetBrains classified roughly 31% of respondents as “agentic coders,” whose output averaged 84% fully agent-generated code.
Seniority adds an interesting twist. Senior developers were slightly more likely than junior developers to fall into that heavy-agent group. Deep AI adoption therefore cannot be explained simply as inexperienced programmers outsourcing work they cannot do themselves.
The normal workflow is becoming a mix: the developer shapes the task, the AI produces more of the implementation, and the developer spends more time checking, steering and integrating what comes back.
Which coding tasks are people comfortable giving AI today?
AI coding adoption is strongest today in implementation, tests, refactoring, migrations, debugging and code review because developers can check those tasks relatively quickly.
The common thread is a tight feedback loop. An agent can modify a function, compile the project, run the tests, inspect an error and retry. A human can then inspect the diff before anything reaches production.
Real company deployments line up with that pattern. Virgin Atlantic uses Codex to improve test coverage. Ramp applies it to code review. Google has used agents for large code migrations. Faros now finds that roughly 25% of pull requests in the engineering organizations it tracks are reviewed by an AI agent.
Stack Overflow’s 2025 survey showed where confidence drops. Some 76% of respondents said they had no plans to use AI for deployment and monitoring, while 69% said the same about project planning.
The split is pretty intuitive. A broken generated test or bad refactor usually leaves evidence developers can inspect. Production operations and ambiguous planning decisions carry broader consequences and much fuzzier definitions of “correct.”
AI agents are spreading fastest through work where software can help verify software.
If you want more recent data on this point, please see our latest AI code assistant market report.
This chart, featured in our AI code assistant market deck, compares the main business model options for AI developer tools platforms
Is AI coding actually making developers faster?
AI coding is probably making many developers faster now, although we still do not have a credible universal productivity number.
METR produced the most useful reality check on this question. Its early-2025 randomized experiment followed 16 experienced open-source developers completing 246 real tasks in mature repositories. The developers expected AI to make them 24% faster. Instead, task completion became 19% slower.
Later results changed direction. When METR tried to repeat the experiment with newer AI tools, original participants showed an estimated 18% speedup and newly recruited developers an estimated 4% speedup. The confidence intervals were wide, and METR explicitly warned that selection effects made the later numbers weak evidence.
The experiment also became harder for an interesting reason. Some experienced developers were increasingly unwilling to participate if the study required them to work without AI. METR found its time measurements becoming less reliable too, because some participants ran several agents concurrently.
Broader surveys point toward real perceived gains. Google’s DORA research found more than 80% of respondents reporting higher productivity with AI. Stack Overflow found 69% of agent users saying agents saved time on specific development tasks.
We can be fairly confident about direction and much less confident about magnitude. AI coding is useful enough that many developers dislike giving it up. Claims that it makes every engineer 30%, 50% or 100% faster remain much harder to defend.
Why do developers keep using AI coding tools when they do not trust them?
Developers keep using AI coding tools because imperfect code can still save time when checking and fixing it takes less effort than writing the first version manually.
Stack Overflow’s 2025 survey captures the contradiction nicely. Some 46% of developers distrusted the accuracy of AI tools, versus 33% who trusted them. Only 3% reported very high trust.
The frustration is equally clear. Two-thirds complained about AI answers that were “almost right,” and 45% said debugging AI-generated code could take more time.
Yet adoption kept climbing after that survey. JetBrains’ more recent numbers show daily coding-agent use at 68% among professional developers.
Developers seem to be applying a practical standard rather than a trust standard. The agent does not need unsupervised control. It needs to make the first draft, search, rewrite or debugging loop cheap enough that supervising the work still feels worthwhile.
That is why trust and adoption can move in opposite directions. Developers can be skeptical about the output and still feel considerably slower without the tool.
This chart, featured in our AI code assistant market deck, illustrates how market revenue is distributed across customer segments in the AI code assistant market
Has AI coding made software teams faster, or just made more code?
AI coding is producing more engineering output now, but review and production systems are struggling to absorb it cleanly.
This is one of the biggest tensions in the market because individual developer speed and team-level delivery speed are different things.
Faros AI’s 2026 report looked at two years of engineering telemetry covering 22,000 developers across more than 4,000 teams. As AI adoption increased, completed tasks per developer rose 33.7%, epics completed rose 66.2% and pull-request merge rate increased 16.2%.
The downstream numbers were much uglier. Pull requests became 51% larger. Bugs per developer rose 54%. Median review time increased more than fivefold, incidents per pull request more than tripled, and code churn increased by more than eight times.
Faros sells engineering-management software, so its interpretation deserves the usual vendor caution. The underlying dataset is still useful because it measures activity in development systems instead of asking developers how productive they feel.
Google’s DORA research found a softer version of the same tension. Higher AI adoption correlated with greater software-delivery throughput, while delivery stability still deteriorated.
The bottleneck is moving. Code has become much cheaper to produce, which places more pressure on review, testing, architecture and production controls.
| What changes with heavier AI use | Recent evidence |
|---|---|
| Tasks completed per developer | +33.7% |
| Epics completed per developer | +66.2% |
| Pull-request size | +51% |
| Bugs per developer | +54% |
| Median pull-request review time | More than 5x |
| Incidents per pull request | More than 3x |
| Code churn | More than 8x |
If you want more recent data on this point, please see our latest AI code assistant market report.
Is “vibe coding” turning into a real business?
AI app building outside traditional engineering has become a real business, with platforms such as Lovable and Replit reaching commercial scale that would have looked implausible two years ago.
Replit gives us one of the clearest trajectories. The company said annualized revenue increased from $2.8 million to $150 million in less than a year, driven by a community of more than 40 million users. Customers include teams at companies such as Zillow, Duolingo and Coinbase.
The important change is who can now create software. AI app builders let founders, designers, product managers and other non-specialists describe an application, generate its interface and logic, iterate through conversation and deploy without assembling the traditional development toolchain themselves.
This creates a different market from Claude Code or Codex. A senior engineer modifying a large production repository and a founder generating a small SaaS product through prompts should not be counted as the same kind of adoption.
They do share one big behavioral shift: writing syntax is becoming a smaller part of creating software. The market for AI coding is expanding beyond people who previously thought of themselves as coders.
This chart, featured in our AI code assistant market deck, shows how AI coding assistant technology has evolved over time
Can AI coding agents actually replace software engineers?
AI coding agents currently replace a growing amount of software-engineering execution, while engineers still hold most of the responsibility for deciding, checking and shipping.
The distinction becomes obvious when we look at where adoption falls off.
Developers are already comfortable delegating implementation, tests, refactoring and parts of review. Some experienced users run several agents at once. Google engineers are orchestrating groups of agents, and the company says three quarters of its new code is now generated by AI before human approval.
Once the task reaches architecture, unclear product requirements or risky production changes, human involvement becomes much harder to remove. Stack Overflow found strong resistance to delegating deployment, monitoring and project planning. Faros’ telemetry also shows what happens when generated output outruns verification: review time, bugs, incidents and rework rise sharply.
The job itself is already changing. A developer who once spent an afternoon manually writing an implementation may now describe the task, inspect the agent’s attempt, correct its direction and decide whether the result is safe to merge.
That is a meaningful substitution of labor. It has arrived first inside the engineer’s job rather than through wholesale replacement of the engineer.
So what is actually getting real adoption in AI coding now?
Agentic coding with human supervision is the clear center of AI coding adoption today.
The strongest evidence now comes from several independent directions. Professional developers use coding agents constantly. Claude Code has become the most common primary choice in JetBrains’ latest survey. Codex grew more than fivefold in professional penetration within roughly half a year. GitHub Copilot has reached 50 million users, its revenue recently accelerated more than 60% quarter over quarter, and one in three GitHub pull requests now involves an agent.
The depth of use has changed too. Developers told JetBrains that agents fully generated roughly 47% of their code on average, while around 22% already delegated more than four-fifths of their coding. As seen above, Google has pushed the model much further internally, with AI generating 75% of new code before engineers approve it.
Companies are paying for this behavior rather than merely experimenting with free tools. Claude Code has passed a $2.5 billion annualized revenue run rate and gets more than half its revenue from enterprise customers. Cursor has reached roughly $1 billion in annualized revenue. Codex has millions of weekly users and is spreading through engineering workflows at large companies.
The remaining constraint has become clearer at the same time. Software teams can generate code faster than they can always review, test and safely absorb it. Faros measured substantially higher throughput alongside more bugs, incidents, review time and churn. Developers themselves remain skeptical of AI accuracy even while using the tools every day.
The part that has genuinely arrived is straightforward: developers describe more of the work, coding agents execute more of it, and humans increasingly spend their time steering, reviewing and deciding what ships.
Full autonomous software engineering still sits further out. The interesting question now is how quickly AI can take over the verification work that currently limits everything it has already learned to generate.
If you want more recent data on this point, please see our latest AI code assistant market report.
In our AI code assistant market deck, we identify pain points entrepreneurs should prioritize
OUR METHODOLOGY
This analysis tests what is actually getting real adoption in AI coding by separating broad AI usage from repeated professional use, deep code delegation, enterprise purchasing, real engineering workflows, measured productivity and downstream software-delivery effects.
We treated developer surveys as the clearest evidence for usage frequency, primary-tool preference and delegation depth. JetBrains’ 2026 research provides the main professional-developer comparison across Claude Code, GitHub Copilot, Codex and Cursor, while Stack Overflow’s 2025 survey gives an earlier benchmark for agent adoption, trust and the tasks developers are reluctant to delegate.
Vendor disclosures were used for absolute scale, revenue and enterprise expansion rather than for direct market-share comparisons. Microsoft’s GitHub Copilot figures, Anthropic’s Claude Code metrics and OpenAI’s Codex disclosures all use different definitions of users and activity, so we kept those measures separate from survey-based professional adoption percentages.
We also separated broad adoption from frontier adoption. Google’s internal figure that AI generates 75% of new code before engineer approval shows how far agentic development can already go inside an unusually advanced engineering organization; it is used here as a marker of depth, not as a benchmark for the average company.
Productivity was assessed with more caution than usage. METR’s controlled developer experiments were used to test direct task-speed claims, while Google’s DORA research and Faros AI’s engineering telemetry were used to examine broader developer productivity, delivery throughput, review pressure, bugs, incidents and code churn.
Where evidence changed quickly, we prioritized the freshest comparable measurement and used earlier surveys to establish direction without pretending differently designed studies form a perfect time series. This is especially important for coding agents, where developer behavior has moved materially within months.
Key sources used for this analysis include JetBrains on coding-agent adoption, JetBrains on professional AI coding tool usage, JetBrains on code delegation, Stack Overflow’s 2025 Developer Survey, Google Cloud’s 2025 DORA report, Anthropic’s Claude Code commercial metrics, OpenAI on scaling Codex to enterprises, Microsoft’s FY2026 Q4 earnings disclosure, Google on internal AI-generated code and code migration, METR’s original productivity experiment, METR’s follow-up experiment update, and Faros AI’s engineering telemetry study.
This chart, featured in our AI code assistant market deck, illustrates how revenue is distributed geographically across Europe, Asia, North America, Africa, and South America in the AI code assistant market
Related blog posts
- AI Coding: what is actually working now?
- The most recent funding news in the AI code assistant market
- AI Coding: what are the top startups?
- What are the key fundraising trends in the AI code assistant market?
- How strong is fundraising in the AI code assistant market right now?
- How funding activity has changed in the AI code assistant market
