AI Coding Assistants Reshuffled: Claude Code 46% Overtakes Cursor 19% as CLI Tools Take the Lead

Forum topic · 小凯 · 2026-08-16

Summary

A JetBrains developer survey from April 2026, widely cited in Chinese AI media, found 46% of engineers named Claude Code their favorite AI coding tool, versus 19% for Cursor and 9% for GitHub Copilot. This forum post contextualizes the numbers: the 46% is a 'favorite' vote, not usage—Stack Overflow's April 2026 Pulse Survey shows Copilot leading usage at 41% with Claude Code at 26%. The post argues a 'mindshare flip' is underway, backed by SitePoint benchmarks showing Claude Code ahead on multi-file reasoning, defect diagnosis, and architectural decisions. It contrasts Cursor's IDE-centric workflow with Claude Code's CLI-based, whole-project reasoning approach, and notes Cursor's continued commercial dominance (95% enterprise AI-vendor purchases per Ramp data). The post also covers a Hacker News-reported 'auto-research' case where developer Sankalp used Codex with Modal to accelerate a GPU QR-decomposition kernel 232x over 14 days and 1,500+ submissions, finishing 12th of 183. It closes with open questions: Anthropic's canary deployments using Claude Code on AI-generated PRs, Cursor's reported $60B acquisition, and whether CLI tools will be absorbed back into IDEs by 2027.

JetBrains' April 2026 developer survey, repeatedly cited by Chinese AI media over the past week, is notable not because it sounds impressive, but because it is irreconcilable: 46% of engineers named Claude Code their favorite AI coding tool, versus 19% for Cursor and 9% for GitHub Copilot. Claude Code's 46% is more than double Cursor's and over five times Copilot's. Claude Code is not an IDE—no autocomplete, no GUI, just a command line. How did it top a field crowded with IDEs?

Putting the Numbers in Context

The 46% in the JetBrains report is a "favorite" vote—a multi-select question, not a usage rate. Cross-referencing with Stack Overflow's April 2026 Pulse Survey (1,100 developers, multi-select, published 2026-05-27):

| Tool | Stack Overflow April Pulse usage (past year) | |---|---| | GitHub Copilot | 41% | | Claude Code | 26% | | OpenAI Codex | 22% | | Cursor | 22% | | Gemini / Code Assist | 21% | | Google Antigravity | 18% | | JetBrains AI | 14% |

Pragmatic Engineer's February 2026 survey of 906 engineers—reported as the original source of the JetBrains figures—found senior engineers' "favorite" at 46% Claude Code vs 9% Copilot. JetBrains AI Pulse January 2026 data meanwhile shows Copilot leading在职 usage at 29% vs Claude Code's 18%—Copilot still leads usage, but Claude Code has flipped the "mindshare."

This mindshare flip is structural, not just sentiment: SitePoint benchmarks show Claude Code consistently leading Cursor on multi-file reasoning, defect diagnosis, and architectural decisions—exactly the capabilities senior engineers value most. Cursor's strength is "autocomplete + IDE integration"; Claude Code's is "read the whole project, plan, modify, verify."

Why the IDE Isn't the End State

Cursor's story: make AI a built-in feature of the IDE. It assumes developers don't need to switch context—everything lives in the IDE.
Claude Code's story: make the IDE one tool of the AI. It assumes developers do switch context—and Claude Code itself is the destination.

When engineers tackle modernization of 50k–500k-line legacy codebases, cross-service refactors, and cross-team PR reviews, their workflow is "deep reasoning + entering unknown code + producing plans," not "autocomplete while typing." Claude Code became the entry point for that workflow.

On the commercial side, Cursor's own June 9, 2026 press release claimed 67% of Fortune 500 companies use Cursor with 150M+ enterprise code lines per day; its updated June 13 enterprise page lowered these to 64% and 100M without explanation. Ramp's June 2026 data: 95% of enterprises that bought an AI-code vendor bought Cursor—still #1 in enterprise procurement. But in the JetBrains "favorite" vote, Cursor fell to 19%: that is individual developer preference.

The Other Story: Codex 232× Kernel Acceleration

An August 15 Hacker News auto-research retrospective offers a different answer. Developer Sankalp spent 14 days using Codex + ChatGPT Pro + Claude Pro + Modal, with 1,500+ submissions, compressing a batched square compact-Householder QR decomposition GPU kernel from 419,000 µs (~419 ms) to 1,805 µs—a 232× speedup.

Caveats:

This was a GPU Mode + Core Automation auto-research competition: fixed problem (qr_v2), automatic scoring, near-unlimited submissions.
Sankalp made 1,500 submissions and kept 560 variants; ranked 12 / 183—2nd place hit 1,228 µs (≈341×), 1st place above 340×. "232×" is not optimal.
The bottleneck was launch overhead and panel overhead, not memory or compute; Sankalp used block Householder + trailing WY update to compress serial work into one column panel, with tensor-core matmul for the rest.
The reproducible method: 3 files (problem_statement.md, AGENTS.md, log.md) + 2 commands (popcorn CLI for submission/benchmark, Modal for profiling).

The core insight: agents crave tight feedback loops. The competition provided an oracle (automatic scoring and verification), a single file, clear boundaries, and near-unlimited submissions—perfect conditions for automated hill-climbing. The three prerequisites: an oracle must exist, verification must be cheap, and results must be quantifiable. Otherwise the AI loop is "accelerating lies toward the wrong target."

Why Both Stories in the Same Week Matter

JetBrains AI Pulse January 2026: 85% of developers use AI tools daily. The JetBrains report says the tools contest is decided—CLI tools took mindshare first place. The Codex 232× case says the winning usage pattern is converting single-point capability into closed-loop iteration. Meanwhile, Microsoft reportedly switched Copilot's default engine to Polaris (MoE 137B / 5B active / self-reported 51.2% on SWE-Bench Pro) on its Maia 200 chip—a giant's counterattack aimed at the enterprise procurement position (where Cursor holds 95%), not developer mindshare, which has already shifted to the "CLI + closed loop" combo.

Three Open Questions

1. Anthropic's own canary deployments use Claude Code heavily to modify AI-generated PRs. Per Anthropic's August 13 research on emerging multi-agent failure modes, AI editing AI-written code can progressively amplify error structures—an industry-wide "AI fixing AI's holes" loop risk. 2. Cursor's reported $60B acquisition (August 14) may reshape its product cadence; if subscription pricing shifts or built-in model weighting changes, the 46% / 19% figures could be rewritten. 3. Will CLI tools be absorbed back into IDEs by 2027? If Anthropic ships a Claude Code for VS Code, today's "CLI vs IDE" contest may become an "Anthropic IDE vs Cursor IDE" contest.

Where This Leaves Us

Claude Code's 46% is individual developer preference; Cursor's 95% is enterprise procurement. Cursor holds the enterprise, Claude Code holds the individual—that is the real split in mid-2026. If enterprise renewals and internal AI-team push keep Cursor's pipeline intact, Cursor still wins commercially; if individual preference propagates into procurement decisions, Cursor's position will be challenged.

Sources

1. JetBrains April 2026 developer survey (cross-validated by Pragmatic Engineer's 906-engineer survey) 2. Stack Overflow April 2026 Pulse Survey (1,100 developers, published 2026-05-27) 3. Cursor June 9, 2026 press release vs June 13 enterprise page data 4. SitePoint benchmarks: Claude Code ahead of Cursor on multi-file reasoning / defect diagnosis / architectural decisions 5. Linxi News / ideaverse.ai auto-research coverage: Sankalp's 12/183 rank and 232× speedup 6. Hacker News "Codex auto-research kernel 232x" thread and Sankalp's retrospective 7. tools.cooconsbit.com / Dev Breakfast 2026-08-16: Codex auto-research analysis 8. 163.com / 硅屿手记: 2026 AI coding assistant reshuffle (2026-08-16)