On 24 February 2025 Anthropic released Claude 3.7 Sonnet and, in the same post, a “limited research preview” of a command-line tool called Claude Code. The stated purpose was modest: “Our goal with Claude Code is to better understand how developers use Claude for coding to inform future model improvements.”
Nobody treated it as the start of a race. But within four months OpenAI had a terminal agent called Codex CLI and Google had Gemini CLI, and the field filled up after that. Now the surveys are in, and one tool stands well ahead of the others.
I want to look at why, because the usual answer is “Anthropic focused on coding.” That is true, and it is not enough. OpenAI also built coding models and a coding agent. What Anthropic did differently is narrower and more useful to copy.
Who launched what
| Date | Tool | What happened |
|---|---|---|
| 24 Feb 2025 | Claude Code | Limited research preview, announced with Claude 3.7 Sonnet |
| 16 Apr 2025 | Codex CLI | Open-source terminal agent from OpenAI, announced as “a new experiment” |
| May 2025 | Claude Code | Made available to the general public |
| 25 Jun 2025 | Gemini CLI | Open-source terminal agent from Google, with a free tier of 1,000 requests a day |
| Aug 2025 | Cursor CLI | Cursor’s agent moved to the terminal |
Plenty of others arrived on the same road: OpenCode, Amp, Kiro, Qwen Code, Copilot CLI. The point of the table is that Claude Code was first by about seven weeks, not by two years. A head start that short does not explain the gap that followed.
Where the race stands
Stack Overflow’s 2026 survey asked 12,255 developers which coding agents they use. Claude Code leads.
| Measure | Value |
|---|---|
| Claude Code | 65.5% |
| GitHub Copilot | 58.7% |
| OpenAI Codex | 29.5% |
| Cursor | 25.5% |
| Google Antigravity | 16.0% |
| Gemini Code Assist | 14.6% |
| OpenCode | 13.6% |
Source: Stack Overflow Developer Survey 2026, AI coding agents. Self-selected respondents; multiple answers allowed.
JetBrains measured something harder to inflate: what developers use at work, weighted to reflect the developer population. In May to July 2026 about 39% used Claude Code, up from 18% in January, and JetBrains says it is “used twice as often as GitHub Copilot”. Copilot fell from 29% a year earlier to 21%.
| Measure | January 2026 | May to July 2026 |
|---|---|---|
| Claude Code | 18% | 39% |
| OpenAI Codex | 3% | 16% |
| Cursor | 18% | 12% |
Source: JetBrains Research, AI coding agents: adoption trends (18 August 2026). January 2026 against May to July 2026.
Anthropic’s own numbers point the same way. In February it said Claude Code’s run-rate revenue had passed $2.5 billion, more than double the figure at the start of 2026, and that weekly active users had doubled since 1 January. These are the company’s figures, not an audit, but the independent surveys agree on the direction.
Open-source mindshare is a different story. OpenCode has 213,866 stars on GitHub, more than Codex (129,381) or Claude Code’s repository (151,044). Only 13.6% of Stack Overflow respondents use it. Stars measure attention. They did not predict use here.
Why it led: a loop, not a niche
Boris Cherny, who created Claude Code, told the Latent Space podcast: “There was no master plan.” He had been experimenting with the public API in different places. Latent Space’s own write-up calls “do the simple thing first” an Anthropic product principle, and points to memory as the example: a markdown file that gets loaded automatically.
That small start mattered because of what it fed. The 3.7 post states the loop plainly: use the tool to learn how developers use the model, then improve the model. I read three parts in it, and each shows up in public sources.
- The people who train the model use the product. The goal line in the 3.7 post is about feeding what developers do back into the next model. That is a loop OpenAI and Google can also build, and both ship strong coding models. I could not find public evidence that either closed it as tightly. Treat that as my reading, not a finding.
- A thin product on a strong model. The Pragmatic Engineer quotes the team: “Every time there’s a new model release, we delete a bunch of code.” With the 4.0 models they deleted around half the system prompt. The harness gets thinner as the model improves, which keeps it quick to change.
- Conventions other people adopt. CLAUDE.md, hooks, subagents and skills are Claude Code features. The Agent Skills format, which Anthropic developed and released as an open standard, is now listed as supported by ChatGPT and Codex, Gemini CLI, Cursor, GitHub Copilot, OpenCode and Kiro. OpenAI ships a plugin whose README says it is “for Claude Code users who want an easy way to start using Codex from the workflow they already have.” When a rival writes a plugin for your product, you are the default.

The lead did not come from doing only coding. It came from using the product to improve the model that powers the product.
Where the lead is cracking
It would be easy to stop there. Four facts argue against the tidy version.
- Codex is closing fast. JetBrains has Codex growing about 5x in at-work use in six months.
- The harness is a smaller edge than the fans say. DeepSeek’s V4.1-Flash card runs one model in eight harnesses. Claude Code scores 69.8 on DeepSWE and 88.0 on Terminal-Bench 2.1. A minimal harness called mini-SWE scores 74.2 and 90.3, ahead of Claude Code on both. I make the same point in the leaderboard post.
- It has stumbled on its own. On 23 April 2026 Anthropic published a postmortem tracing reports of worse quality to three changes in Claude Code and its SDK. “The API was not impacted.” The model was fine. The harness was the cause, which shows how much weight the harness now carries.
- Part of the lead may be a subscription effect. Claude’s docs say subscription sign-in is “intended exclusively for purchasers of Claude Free, Pro, Max, Team, and Enterprise subscription plans” and for “ordinary use of Claude Code and other native Anthropic applications.” If flat-rate plans pull developers into Claude Code, some of the share is price, not product.

Our take
The lesson I take is about loops. Focus on coding was necessary, since every serious lab has it. What set Claude Code apart was a tight cycle between a model team, a thin harness and daily users, plus conventions that others copied.
Three things would change my mind:
- Codex passes Claude Code in the next independent survey wave while Claude keeps the best coding model. That would say the product loop matters less than the model.
- Claude Code’s share among developers who pay by the API token, not a subscription, is no higher than Codex’s. That would say the plans are doing the work.
- A harness that works with every model wins share. OpenCode has 42% awareness and 7% use at work, so the demand exists and has not converted.
If you are building an agent product, put your own builders inside it every day, keep it thinner than your model, and publish the conventions you want others to copy. If you are choosing a tool, treat the harness as a variable and test two of them on your own work.
References
- Claude 3.7 Sonnet and Claude Code Anthropic, 24 February 2025.
- Anthropic raises $30 billion in Series G funding Anthropic, 12 February 2026.
- Stack Overflow Developer Survey 2026: AI coding agents Stack Overflow, 2026.
- AI coding agents: adoption trends JetBrains Research, 18 August 2026.
- Claude Code (Latent Space interview with Boris Cherny) Latent Space, 7 May 2025.
- How Claude Code is built The Pragmatic Engineer, 23 September 2025.
- codex-plugin-cc OpenAI on GitHub.
- Gemini CLI: your open-source AI agent Google, 25 June 2025.
- An update on recent Claude Code quality reports Anthropic Engineering, 23 April 2026.
- Authentication and credential use Claude Code docs.
- DeepSeek-V4.1-Flash model card DeepSeek, Hugging Face.
- Agent Skills agentskills.io.