Key findings
- Claude Code and GitHub Copilot lead reported use. Of 12,255 Stack Overflow 2026 respondents, 65.5% used Claude Code and 58.7% used Copilot in the past year, ahead of OpenAI Codex at 29.5% and Cursor at 25.5%.
- Use shifted within months. JetBrains measured Claude Code at work rising from 18% of professional developers in January 2026 to 39% in May to July, and Codex from 3% to 16%. GitHub Copilot fell to 21%, from 29% a year earlier.
- On METR's task suite, the length of task a frontier model completes half the time has doubled about every 129 days since 2023. It grew from 4 minutes for GPT-4 to about 12 hours for Claude Opus 4.6 in February 2026.
- Trust stays conditional. Only 6.6% of Stack Overflow 2026 respondents trust AI output with important work decisions, and 76.5% run AI output locally before they use it.
Claude Code leads coding agent use at work in 2026, and engineers still check agent output by hand before they trust it. Agents are now the main way developers use AI at work. In Stack Overflow's 2026 survey, 65.9% of respondents use a coding assistant or agent at work, and JetBrains finds 90% of professional developers use agents at least weekly. Write your review and permission rules now, because they apply whichever agent your engineers pick next quarter.
Who uses which agent
Claude Code and GitHub Copilot sit far ahead of the rest, with Cursor and Codex in a second group. Stack Overflow asked respondents which coding agents or assistants they had used over the past year. Most named more than one, so the shares add up to well over 100%.
Stack Overflow draws its sample from its own community and does not reweight it, so its shares measure reach among engaged developers. JetBrains reweights its survey to the global developer population and finds the same leader. In May to July 2026, 39% of professional developers used Claude Code at work, and 47% in the United States.
| Product | Share of respondents |
|---|---|
| Claude Code | |
| GitHub Copilot | |
| OpenAI Codex | |
| Cursor | |
| Google Antigravity | |
| Gemini Code Assist | |
| Windsurf (now Devin Desktop) | |
| Cline | |
| Devin |
Stack Overflow Developer Survey 2026, n = 12,255 respondents to the coding agent question, fielded June 23 to August 5, 2026, published October 6, 2026. Multiple answers allowed.
How fast the market moved in 2026
Claude Code more than doubled its share of developers using it at work between January and mid-2026, and Codex grew fivefold. Cursor and GitHub Copilot lost share. JetBrains asked the same question twice, in its January AI Pulse survey and in its Developer Ecosystem Survey from May to July.
Two products changed shape during the year. On June 18, 2026, Gemini CLI stopped serving Google AI Pro and Ultra subscribers and free Gemini Code Assist users, who move to Antigravity CLI. Gemini Code Assist Standard and Enterprise customers keep their access. Windsurf became Devin Desktop on June 2, 2026, with plans and pricing unchanged, and Devin's Pro, Max, and Teams full seats now include it.
| Product | January 2026 | May to July 2026 |
|---|---|---|
| Claude Code | 18% | |
| GitHub Copilot | 29% | |
| Codex | 3% | |
| Cursor | 18% | |
| JetBrains AI (AI Assistant and Junie) | 11% | |
| OpenCode | ||
| Google Antigravity | 6% |
JetBrains Research. January 2026: AI Pulse, 10,000+ professional developers. May to July 2026: Developer Ecosystem Survey 2026, 15,000+ professional developers, reweighted by region, employment, language, and familiarity with JetBrains products.
How much work an agent finishes on its own
A model's time horizon, in METR's measure, is the length of task it completes half the time, with each task timed by how long human experts take. On version 1.1 of METR's task suite the frontier horizon has doubled roughly every 129 days since 2023 (95% interval 104 to 158 days). Claude Opus 4.6 reached about 12 hours in February 2026, with a wide interval of 5.3 to 60.6 hours.
The newest models have outgrown the suite, and METR now reports them in its evaluation write-ups. METR states that measurements above 16 hours are unreliable on the current suite, and it stopped updating its horizon page on September 8, 2026. In its May 2026 Frontier Risk Report, the most capable model shared with METR had a point estimate past that limit and below 20 hours. For GPT-5.6 Sol in June 2026, METR's standard method gives about 11.3 hours (95% interval 5 to 40 hours). METR does not treat that figure as a reliable measurement, because the model cheated on tasks more often than any public model it had tested. For Claude Opus 5.5 in September 2026, METR used five AI research tasks in place of a time horizon.
A horizon counts tasks that pass automated checks, which is a lower bar than review. In METR's March 2026 review, repository maintainers would not have merged roughly half of the test-passing pull requests agents wrote.
| Model | Release | Time horizon |
|---|---|---|
| GPT-4 | Mar 2023 | |
| Claude 3.7 Sonnet | Feb 2025 | |
| o3 | Apr 2025 | |
| GPT-5 | Aug 2025 | |
| Claude Opus 4.5 | Nov 2025 | |
| GPT-5.2 | Dec 2025 | |
| Claude Opus 4.6 | Feb 2026 | |
| Claude Mythos Preview (early) | Apr 2026 |
METR Time Horizon 1.1 benchmark results, frontier models only, as of METR's last page update. METR treats values above 960 minutes as unreliable.
What engineers still check by hand
Developers trust agents with work they can verify. In Stack Overflow's 2026 survey, 48.0% of 14,304 respondents trust AI output when they can easily verify it, 16.3% only for low-risk tasks, and 6.6% for important work decisions. Asked how well AI handles complex tasks, 16.0% said very well and 38.8% good but not great.
Agents already write a large share of code. JetBrains' 2026 survey estimates that developers have agents write about 47% of their code outright, and that about 22% of developers rely on agents for more than 80% of it. The larger that share, the more each of the checks below matters.
| Step | Share of respondents |
|---|---|
| Run it locally | |
| Compare against codebase context | |
| Inspect tests and security implications | |
| Read official documentation | |
| Search Stack Overflow | |
| Use it as-is |
Stack Overflow Developer Survey 2026, n = 13,162. Multiple answers allowed.
How organizations govern agent use
Most organizations have not set a policy. In Stack Overflow's 2026 survey, 29.7% of respondents say AI use is optional and left to individuals, 24.0% work with approved tools and written guidance. Also, 10.2% are measured on AI use in performance reviews. On tooling, 32.8% use a few approved tools and 18.6% mostly one.
An organization without approved tools still runs agents, under each engineer's personal plan and data terms. Decide which plans your engineers use, which permissions agents get by default, and who reviews agent output before usage grows further.
| How the organization treats AI | Share of respondents |
|---|---|
| Optional, left to individuals | |
| Approved tools and guidance | |
| Encouraged by some teams, practices vary | |
| Expected for some types of work | |
| Measured in performance reviews | |
| Not used for work | |
| Not sure |
Stack Overflow Developer Survey 2026, n = 13,167.
Methodology
Survey figures come from the Stack Overflow Developer Survey 2026 (30,903 valid responses, fielded June 23 to August 5, 2026) and JetBrains Research's 2026 survey posts, each with its n and window. Stack Overflow's respondents select themselves from its community, and JetBrains reweights its sample. Time horizons come from METR's Time Horizon 1.1 data and its 2026 evaluation reports, and product changes from the vendors' own pages. Every figure is its publisher's own, linked under Sources.
Products in this piece
- Claude Code+75 Strong positiveAI Coding Agent by Anthropic
- Codex+55 Strong positiveAI Coding Agent by OpenAI
- Cursor+11 MixedAI Coding Agent by Anysphere
- GitHub Copilot+15 MixedAI Coding Agent by GitHub
- Gemini CLI−3 MixedAI Coding Agent by Google
- Devin+39 PositiveAI Coding Agent by Cognition
- Devin Desktop (formerly Windsurf)+17 MixedAI Coding Agent, also by Cognition
- Cline+39 PositiveAI Coding Agent by Cline
- Junie+28 PositiveAI Coding Agent by JetBrains
Sources
- Public source
- 12
- Official documentation
- 4
- Stack Overflow: The results of the 2026 Developer Survey are here (Oct 6, 2026)
- Stack Overflow Developer Survey 2026: Methodology (30,903 responses, fielded Jun 23 to Aug 5, 2026)
- Stack Overflow Developer Survey 2026: AI, full data (published Oct 6, 2026)
- JetBrains Research: AI Coding Agents, Adoption Trends (Developer Ecosystem Survey 2026, Aug 2026)
- JetBrains Research: Which AI Coding Tools Do Developers Actually Use at Work? (AI Pulse, Apr 2026)
- JetBrains Research: How Much Code Do Developers Really Let Agents Write? (Aug 2026)
- METR: Task-Completion Time Horizons of Frontier AI Models (Time Horizon 1.1, last updated Sep 8, 2026)
- METR: Time Horizon 1.1 benchmark results (data file, last updated Sep 8, 2026)
- METR: Many SWE-bench-Passing PRs Would Not Be Merged into Main (Mar 10, 2026)
- METR: Frontier Risk Report, February to March 2026 (May 19, 2026)
- METR: Summary of METR's predeployment evaluation of GPT-5.6 Sol (Jun 26, 2026)
- METR: Summary of METR's predeployment evaluation of Claude Opus 5.5 (Sep 22, 2026)
- Google Developers Blog: Transitioning Gemini CLI to Antigravity CLI (May 19, 2026)
- Gemini CLI docs: Quota and pricing
- Devin docs: Self-serve plans
- Devin Desktop docs: FAQ on the move from Windsurf