Claude Code is winning the terminal, and the numbers from April 2026 are not subtle

On April 14, 2026, Anthropic shipped Routines. On April 15, Google announced sub-agents had finally arrived in Gemini CLI. On April 16, Claude Opus 4.7 hit 87.6% on SWE-bench Verified. On April 23, OpenAI shipped GPT-5.5 with a state-of-the-art Terminal-Bench 2.0 score.

Between April 20 and 24, Codex CLI rolled out plugin management, Plan Mode, and stable hooks across versions 0.122 through 0.125. One week in April 2026, five releases, one verdict: in the terminal, the rest are still chasing primitives Claude Code already shipped.

I have not opened Codex CLI or Gemini CLI in three months. After looking at April 2026's benchmarks alongside the latest survey data and ecosystem counts, I am not planning to. Here is the bold claim I am making and earning across the rest of this piece.

Claude Code does not just lead the terminal AI category in April 2026. It defines it. Codex CLI is OpenAI's catch-up answer. Gemini CLI is Google's. Both have closed real ground in 2026. Neither has closed the structural gap, and the structural gap is the ecosystem, the workflow primitives, and roughly 9 months between when Claude Code shipped sub-agents, hooks, plugins, and skills, and when Gemini caught up to one of those.

The moat is not the model. It is developer mindshare and skill-catalog inertia. A 46% most-loved tool in the Pragmatic Engineer 2026 survey of 906 developers, sitting on top of a 4,200-skill catalog as of April 28, 2026, is not something competitors close in a quarterly release.

Across four dimensions, ecosystem, output quality on real coding tasks, terminal-native UX, and market position, the numbers point one way. I will show you which numbers and where they come from.

The four dimensions, with receipts

Ecosystem.

As of April 28, 2026, the Claude Code plugin marketplace tracked 4,200+ skills, 770+ MCP servers, and 2,500+ community marketplaces. Codex CLI's main extension model is third-party MCP shims, things like codex-subagents-mcp, which the community had to build because the first-party feature was missing for over a year. Gemini CLI shipped Agent Skills in March 2026, but the catalog is dwarfed.

Skills, MCPs, sub-agents, hooks, plugins, Routines: this is the vocabulary builders use to assemble systems. Claude Code shipped that vocabulary first, and the others are catching up to it.

Output quality on real coding tasks.

In April 2026 the picture is two benchmarks, two leaders, one ecosystem. Claude Opus 4.7, released April 16, 2026, leads SWE-bench Verified at 87.6% and SWE-bench Pro at 64.3%. SWE-bench Pro is the hard one, real issues from open-source repos.

OpenAI's GPT-5.5, released April 23, 2026, hit 82.7% on Terminal-Bench 2.0, which OpenAI calls state-of-the-art on that benchmark. On SWE-bench Pro, GPT-5.5 sits at 58.6%, behind Opus 4.7. Gemini 3.1 Pro trails both at 54.2%.

Agent scaffolding matters as much as the raw model. The headline I read off this in April 2026: Claude leads on the harder real-world repository benchmark; OpenAI's newer model leads on terminal-task tooling; the ecosystem still favors Claude Code by a wide margin.

Terminal-native UX.

Claude Code runs across 6 surfaces in 2026: CLI, VS Code, JetBrains, a standalone desktop app, claude.ai/code on the web, and iOS. It also ships 1 native cloud-scheduled job primitive: Routines. Codex CLI has 4 surfaces (ChatGPT web, Codex CLI, desktop for Windows and macOS, IDE integrations) and 0 native scheduled-job primitives.

Gemini CLI has 2 surfaces (CLI and a VS Code companion) and 0 native scheduled-job primitives. If you want a Routine equivalent on Codex or Gemini, you bolt on cron, GitHub Actions, or third-party orchestration. That is not the same thing as a first-class primitive that wakes up every six hours, runs your prompt against your repo with your connectors, and posts the result to Slack on failure. The Register, on April 14, 2026, called a Routine "a dynamic cron job or a trigger-driven, short-lived agent." Right now, Claude Code is the only one of the three with that primitive built in.

Claude Code's market position and ecosystem maturity.

Claude Code launched as a research preview in February 2025 and went generally available in May 2025 alongside Claude 4. Codex CLI followed on April 16, 2025. Gemini CLI followed on June 25, 2025. The launch dates closed quickly.

The feature-parity gap did not. Look at one week in April 2026: Anthropic shipped Routines on April 14, Gemini shipped sub-agents on April 15, Codex CLI rolled out plugin management, hooks, and Plan Mode across 0.122 through 0.125 between April 20 and 24, and OpenAI shipped GPT-5.5 on April 23. Anthropic invented the playbook; the others are running it on a one-week delay. According to UncoverAlpha, Claude Code reached the $1B run-rate faster than any other AI coding tool in history, with daily VS Code-tracked installs going from 17.7M to 29M since January 1, 2026.

What this looks like on a real workflow

Numbers in tables are one thing. A clock on a real task is another.

In April 2026, the team at codeant.ai ran the same Express.js refactor across all three CLIs and published the runtimes on April 28. Claude Code: 1 hour 17 minutes, zero developer interventions. Codex CLI: 1 hour 41 minutes.

Gemini CLI: 2 hours 4 minutes, with 3 manual corrections from the developer. Same task, same scope, three CLIs, three clocks. Claude Code finished roughly 23% faster than Codex CLI and 38% faster than Gemini CLI, and did not need a single human nudge. That is what "ecosystem maturity" feels like when the prompt hits the keyboard.

Independent surveys triangulate the same picture. The Pragmatic Engineer 2026 AI Tooling Survey (Gergely Orosz, 906 respondents, fielded January 27 to February 17, 2026, published March 3) found Claude Code "most loved" at 46%, with Cursor at 19% and GitHub Copilot at 9%. At companies under 200 employees, Claude Code adoption hit 75%.

Orosz's own line: "Claude Code has gone from zero to be the #1 tool in only eight months." A separate January 2026 academic study from UC San Diego and Cornell, sample of 99 professional developers, put Claude Code at 58 users, ahead of GitHub Copilot at 53 and Cursor at 51. One industry newsletter, one academic study, same direction.

Money is a vote. Claude Code alone was reportedly contributing about $2.5B run-rate by February 2026, sitting inside Anthropic's roughly $30B annualized revenue in March 2026 (up about 1,400% year over year). Builders are paying with their dollars, not just their tweets.

The honest counter-arguments

Numbers like these usually attract three counter-arguments. Let me take them seriously.

Gemini CLI is free and generous.

Yes, in June 2025 it was. Apache 2.0 license, real free tier, 60 requests per minute, 1,000 requests per day. That was the pitch. On March 18, 2026, Google announced that starting March 25, free users would only access Gemini Flash, not Gemini Pro.

According to community reports in the official GitHub discussion thread, Flash daily quota dropped from roughly 1,000 to roughly 150 requests per day. The thread collected 933 thumbs-down reactions from developers who had built workflows on the original promise. Hobbyists may still squeeze value out of what is left.

Serious builders bump quota walls inside an afternoon. The "free and generous" line stopped being true in late March 2026.

But Codex is from OpenAI, the biggest model lab.

Codex has two real wins. First, token cost: in the nxcode.io Figma-to-code benchmark from March 27, 2026, Codex CLI used roughly 1.5M tokens versus Claude Code's 6.2M for comparable output. Roughly 4x cheaper per task.

That is not vibes. That is a bill, especially for high-volume builders. Second, in a Reddit poll cited in the same nxcode.io report (n=500+), 65% of developers said they preferred Codex for daily coding.

Self-reported preference is real signal, even if it is also vibes. Take both wins seriously.

Now the dismantle. In the same nxcode.io report, blind reviews of the actual code output rated Claude Code best in 67% of cases, Codex in 25%, with 8% ties. When the reviewer cannot see which tool produced the code, Claude wins by a factor of two-and-a-half.

On raw benchmarks, Opus 4.7 leads SWE-bench Pro at 64.3% in April 2026 versus GPT-5.5 (released April 23, 2026) at 58.6%. Yes, GPT-5.5 took the Terminal-Bench 2.0 lead a week ago. That is exactly the kind of move I expect: frontier models leapfrog each other on individual benchmarks within weeks.

What does not leapfrog on a release cadence is the ecosystem the work actually runs on. Anthropic's $30B annualized run-rate in March 2026 says the market is voting with money, even without OpenAI's distribution channel.

Here is the named scenario split, because steelmen deserve clean answers. If you are doing high-volume routine generation where the output is fungible (boilerplate, templated CRUD, throwaway prototypes), Codex's 4x token-cost win is real. Take it.

If you are shipping production code that humans will read, debug, and extend, Claude Code's blind-review win is the one that pays the bigger bill back. Cleaner code costs less in rework. Comparing the two for the work most builders are actually paid to do is the difference between a Formula 1 car and a bicycle.

Both are vehicles. One is built for the work production builders are paid to do.

But the gap is closing. Gemini just shipped sub-agents. Codex shipped Plan Mode. GPT-5.5 just took a benchmark lead.

True on every count. Gemini CLI got native sub-agents on April 15, 2026. Codex CLI shipped plugin management, hooks, and Plan Mode in 0.122 through 0.125 between April 20 and 24, 2026. GPT-5.5 dropped April 23.

These are real catch-up moves, not parity. Claude Code shipped sub-agents in mid-2025, alongside hooks, skills, and the rest of the extensibility stack. The plugin marketplace passed 4,200+ skills as of April 28, 2026.

Routines shipped April 14, 2026, and Opus 4.7 still leads SWE-bench Pro. The rest are not standing still, but neither is Claude Code, and the gap on the primitives that compose into systems is roughly 9 months on the single feature Google just caught up to.

This is what skill-catalog inertia looks like. Skill authors building on Claude Code in April 2026 have little incentive to reauthor for a competitor whose catalog is a fraction of the size, and the surface developers reach for is the surface they keep building on. A 46% most-loved tool sitting on 4,200+ skills is not a thing you close on a release cadence. The honest counter inside this counter is single-vendor lock-in, and the honest answer is that MCPs are vendor-neutral by design (cross-CLI portability is real today) and Routines run server-side, which reduces the terminal-side dependency footprint anyway.

The verdict, and where the rest of the stack fits

Claude Code is the center of a builder's stack in April 2026. Codex CLI is the side dish OpenAI added to keep ChatGPT Pro buyers from looking elsewhere. Gemini CLI is the free-tier sandbox that just got nerfed. Neither is the answer for a serious builder.

The stack I run, and the one I teach in AI Accelerator, looks like this. Claude Code is the brain, the surface where I think and where my sub-agents live. n8n is the nervous system, the workflow automation layer that connects external services and triggers when Claude Code is not the right place for a particular branch. Replicate, accessed through a Claude Code skill, is where I generate images, video, and audio. Whatever the model, Flux for images, ElevenLabs for voice, video gens, I run it from inside Claude Code. Google AI Studio is where I reach when I need a multimodal Gemini-model capability for a specific task, called from inside a Claude Code workflow. OpenClaw is where I run 24/7 autonomous agents that need to work without me sitting at the terminal. MCPs are the bridge layer that pulls every external service inside Claude Code so I never have to leave it.

Together, that is a complete ecosystem. The center holds because Claude Code is the surface that everything else plugs into. Codex CLI does not host an ecosystem like this in April 2026. Gemini CLI is still building one.

Watch the leaderboard or build on top of it. That's the choice in April 2026.