AI Coding Assistant Shake-up: Claude Code Leads 'Favorite Tool' Poll at 46%, Overtaking Cursor (19%) and GitHub Copilot (9%)

Forum topic · 小凯 · 2026-08-16

Summary

A JetBrains developer survey from April 2026, widely cited by Chinese AI media, found that 46% of engineers named Claude Code their favorite AI coding tool—more than double Cursor (19%) and five times GitHub Copilot (9%). This article contextualizes the numbers: the 46% reflects 'favorite tool' preference, not usage share. Stack Overflow's April 2026 Pulse Survey shows usage rates where Copilot leads (41%) ahead of Claude Code (26%), OpenAI Codex (22%), and Cursor (22%), while JetBrains AI Pulse data confirms Copilot still leads workplace usage. The article argues Claude Code's terminal-first design—reading entire projects, planning, editing, verifying—suits senior engineers' workflows for large-scale refactoring and modernization, flipping 'mindshare' even as Cursor dominates enterprise procurement (95% of enterprise AI-coding buyers per Ramp data, with conflicting self-reported Fortune 500 figures). A companion story examines a 14-day Codex-driven auto-research competition entry that achieved a 232x speedup on a GPU QR-decomposition kernel via 1,500+ automated submissions, illustrating how agents thrive in tight feedback loops with cheap verification. The piece concludes that Claude Code holds developer mindshare while Cursor holds enterprise share, and flags open questions around AI-revising-AI code, Cursor's acquisition, and whether CLI tools will be absorbed back into IDEs.

A JetBrains developer survey conducted in April 2026 was repeatedly cited by Chinese AI media this past week—not because it sounds impressive, but because it is irreconcilable with the status quo. In the report, 46% of engineers rated Claude Code as their favorite AI coding tool, versus Cursor at 19% and GitHub Copilot at 9%. Claude Code's 46% is more than double Cursor's and more than five times Copilot's.

Claude Code is not an IDE. No autocomplete, no graphical interface—just a command line. How did it top a track crowded with IDEs?

Putting the Numbers in Context

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

| Tool | Stack Overflow April Pulse usage (used in 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—the original source of the JetBrains figure—found senior engineers' "favorite" tool was Claude Code at 46% vs. Copilot at 9%. Meanwhile, JetBrains AI Pulse data from January 2026 shows Copilot still leading workplace usage at 29% vs. Claude Code at 18%—usage favors Copilot, but mindshare has already flipped.

This mindshare flip is not just sentiment; SitePoint benchmarks reach a structural conclusion: Claude Code consistently leads Cursor in multi-file reasoning, bug 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 Is Not the Endgame

Cursor's story is "making AI a built-in feature of the IDE." Claude Code's story is "making the IDE a tool of the AI." The difference:

Cursor assumes developers don't need to switch context—all AI capability lives inside the IDE.
Claude Code assumes developers do need to switch context—and Claude Code itself is the context you switch to.

The real divide is not "who understands code better" but "who understands developer workflow better." When engineers modernize 50,000–500,000-line legacy codebases, do cross-service refactors, or review PRs across teams, the workflow is "deep reasoning + enter unknown code + produce a plan," not "autocomplete + type-as-you-go." Claude Code has become the entry point for that workflow.

Cursor's own June 9, 2026 press release claimed 67% of Fortune 500 companies use Cursor with 150M+ enterprise code lines daily; its enterprise page, updated June 13, revised that to 64% and 100M lines—without explanation. Ramp data from June 2026: 95% of enterprises that bought an AI-code vendor bought Cursor, ranking it first in the category.

These numbers say "commercial success"—Cursor remains #1 in enterprise procurement. But the JetBrains "favorite" score of 19% shows developer personal preference has been halved by Claude Code.

The Other Story: A 232x Codex Kernel

The JetBrains report answers "whose tool to use," not "how far a tool can be pushed." An August 15 Hacker News auto-research retrospective offers another answer—developer Sankalp used Codex + ChatGPT Pro + Claude Pro + Modal over 14 days, with 1,500+ submissions, compressing a batched square compact-Householder QR decomposition GPU kernel from 419,000 microseconds (~419 ms) to 1,805 microseconds: a 232x speedup.

But 232x needs unpacking:

It was a GPU Mode + Core Automation auto-research contest with a fixed problem (qr_v2), automated scoring (per-shape timings + geometric mean), and near-unlimited submissions.
Sankalp made 1,500 submissions, keeping 560 variants in the final directory.
Rank: 12/183 at 232x—but second place hit 1,228 microseconds (~341x), and first place exceeded 340x. "232x" is not the optimal result; it's what rank 12 achieved.
The bottleneck was not memory or compute but launch overhead and panel overhead—Sankalp used block Householder + trailing WY update to compress serial work into one column panel, leaving the rest to tensor-core matmul.
The reproducible method: 3 files + 2 commands. problem_statement.md + AGENTS.md + log.md; the popcorn CLI (submissions + benchmarking) + Modal (profiling).

At its core: agents crave tight feedback loops. The contest provided an oracle (automatic scoring and verification), a single file, clear boundaries, and near-infinite submissions—perfect conditions for closed-loop, automated hill-climbing.

Sankalp's 14 days and 1,500 iterations equal roughly "4 years of manual human engineering work." The precondition: an oracle must exist, verification must be cheap, results must be quantifiable. Without all three, closed-loop AI becomes "accelerated lying toward the wrong objective."

What It Means That Both Stories Landed the Same Week

Not a coincidence. The JetBrains report says "the AI coding tool race is decided: CLI tools took mindshare first." The Codex 232x story says "the winning pattern for using AI tools: convert single-point capability into closed-loop iteration."

JetBrains AI Pulse (January 2026): 85% of developers use AI tools daily. The Codex 232x demo shows how even the remaining 15% get pulled in: "hand that slow function to Codex for three rounds of profile + patch loops, look at the real speedup, then decide whether to adopt it."

In August, Microsoft switched Copilot's default engine to Polaris (MoE 137B / 5B active / self-reported 51.2% on SWE-Bench Pro) running on its own Maia 200—the giants' counterattack. But the giants' battleground is enterprise procurement (where Cursor already has 95%), not developer mindshare, which has settled on the "CLI + closed loop" combination of Claude Code at 46% and Codex at 232x.

Three Open Questions

1. Anthropic's own canary deployments use Claude Code extensively to modify AI-generated PRs. Per Anthropic's August 13 research on Emerging Multi-Agent Failure Modes, "AI revising AI-written code" gradually amplifies erroneous structure. If Claude Code's 46% mindshare keeps growing, the industry could enter a loop of "AI patching AI's potholes." 2. After Cursor's $60B acquisition by SpaceX (August 14), Cursor's product cadence may be reshaped by SpaceX-AI's internal GPU cost structure—Grok 4.6 is already built into Cursor. If subscription pricing shifts or in-IDE Grok weighting rises, the 46%/19% numbers could be rewritten. 3. Will CLI tools be reabsorbed by IDEs by 2027? If Anthropic ships a Claude Code for VS Code (à la Cursor), JetBrains' 46% could become the win rate of a "Claude Code IDE suite." Today it looks like "CLI vs. IDE"; tomorrow it may just be "Anthropic IDE vs. Cursor IDE."

Where This Leaves Us

Claude Code's 46% is a personal-preference number; Cursor's 95% is an enterprise-procurement number. Cursor owns the enterprise; Claude Code owns the individual—that is the real distribution of AI coding tools in mid-2026. If the enterprise chain of "license renewals + internal AI team mandates" holds, Cursor's business keeps winning through 2026–2027; if developer preference propagates into procurement decisions, Cursor's commercial position faces a challenge.

References

1. JetBrains April 2026 developer survey (cross-validated via 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 leads Cursor in multi-file reasoning / bug diagnosis / architectural decisions 5. Linxi News / ideaverse.ai auto-research coverage: Sankalp's 12/183 rank and 232x 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 explainer 8. 163.com / Silicon Isle Notes: 2026 AI coding assistant reshuffle (2026-08-16)