Three years ago the fight over AI coding tools was a fight over the IDE — whose autocomplete, whose sidebar, whose inline diff. That fight is mostly settled and mostly beside the point now, because the agents that do real work moved into the terminal. Claude Code, OpenAI's Codex CLI, and Google's Gemini CLI are all the same animal: a coding agent that reads your repo, edits files, runs your build and tests, and iterates — in the one place that scripts, pipes, and lives in CI.

They look interchangeable. They are not. Each is a different bet about how you win the terminal, and the cleanest way to choose is to name the bet rather than chase a benchmark.

The closed bet: Claude Code

Claude Code is Anthropic's proprietary agent, tuned tightly around its Claude models. You don't get to read the harness, and there is no free tier — you reach it through a Pro or Max subscription or pay per token. What you get in exchange is the bet that the model and the scaffolding, co-designed and closed, produce the best output on hard work: multi-file refactors, large changes that have to stay coherent across a codebase. It is the option you pick when output quality is the only axis you care about and you're willing to pay and to trust a black box to get it.

The open-harness bet: Codex CLI

OpenAI's Codex CLI makes the opposite trade on openness while keeping a frontier model at the center. The CLI itself is open source — you can read it, extend it, see how it drives the model — wrapped around OpenAI's GPT-5-class models. The bet here is that an open, hackable harness around a top-tier model beats a closed one you can't inspect: you can audit what it does, bend it to your workflow, and you aren't fully stranded if the terms change, because the part you run is yours.

The free-tier bet — and the cautionary tale: Gemini CLI

Google's Gemini CLI entered as the land-grab play: open source, and a genuinely generous free tier that asked for no credit card. For a lot of developers that made it the default for casual and exploratory work — the agent you reach for precisely because it costs nothing.

Then it became the cautionary tale. On May 19, 2026, at Google I/O, Google announced it was consolidating its developer tooling under the Antigravity brand and retiring the standalone Gemini CLI. The hard date is close: on June 18, 2026, Gemini CLI and the Gemini Code Assist IDE extensions stop serving requests for Google AI Pro, Google AI Ultra, and the free Gemini Code Assist tier. The replacement, Antigravity CLI, is rebuilt in Go and open to everyone, and Google says it keeps the load-bearing features — Agent Skills, Hooks, Subagents, and Extensions-now-plugins. Gemini CLI itself doesn't vanish; it survives behind paid and enterprise API keys, as The Register noted with some sharpness about the bait-and-switch optics.

The most useful fact in any 2026 comparison of these three isn't a benchmark score. It's that one of them changed its name and its terms this month — which tells you the axis that actually matters.

That axis is roadmap risk. Benchmarks move every release; this month it's one model's terminal score, next month another's. What doesn't reset is the structural question: closed-and-paid (Claude Code), open-and-paid (Codex CLI), or riding a vendor's free tier that the vendor can re-brand and re-price on a quarter's notice (the Gemini lineage). The free-CLI phase of this market is visibly ending, and the Antigravity transition is the receipt.

How to choose

If you've only ever judged these tools from inside an editor, the same three bets show up there too, drawn differently — that's the IDE-side version of this fight, and it's worth reading next to this one. But the terminal is where the agents actually run now, and in the terminal the choice isn't a leaderboard. It's which company's plan for the next year you're willing to live inside.