If you want a coding agent in your terminal and you don't want to think about a bill, the answer in August 2026 is short: run Codex. OpenAI made it free on every ChatGPT plan, including the $0 tier, and the CLI logs in with your ChatGPT account instead of a metered API key. That's the genuine floor now — no card, no per-token math.

But the floor moved twice this quarter, and the second move is the one that trips people up: the other famous free coding agent, Gemini CLI, is gone. On June 18, 2026 Google stopped serving it for free, AI Pro, and Ultra personal accounts — no grace period — and pushed everyone to a closed-source replacement whose free quota is a fraction of what made the original a 105k-star favorite. If your muscle memory says "just use the free Gemini CLI," that muscle memory is now a broken CI pipeline.

So here's the honest decision for a team of one, four agents deep, with the catch in each.

Codex CLI — the real $0 starting point#

Codex is the cheapest because it's free, full stop. As of 2026 it ships on Free, Go, Plus, Pro, Business, and Enterprise ChatGPT plans, and the CLI authenticates through your ChatGPT sign-in — you don't touch an API key or a usage dashboard to start.

npm i -g @openai/codex     # or: brew install codex
codex                      # sign in with your ChatGPT account

The catch is the meter. On the free plan you get roughly 15–80 local Codex tasks per rolling 5-hour window, with a weekly ceiling on top. That's plenty for writing a script, chasing a bug, or a focused refactor — and not enough for an all-day autonomous agent that churns through a large repo. Cloud features like automated GitHub code review sit behind Plus ($20/mo) or higher. Treat free Codex as the place you prototype and debug, not the place you run production-grade agent volume.

"Free" on a coding agent is a quota, not a capability. The question is never whether it costs nothing — it's what happens at 4pm when the 5-hour window runs dry mid-task.

Kimi Code — the cheapest agent you actually pay for#

When you outgrow Codex's window, the cheapest uninterrupted driver is Moonshot's Kimi Code, which turned on USD tiers starting around $19/mo — and it runs against the open 2.8-trillion-parameter Kimi K3 weights, so the escape hatch is real: if you outgrow the hosted tier, you can point the same agent at the open model. We covered the full decision in Kimi Code vs Claude Code vs Codex CLI, and how to wire a terminal agent to K3 in point your coding agent at Kimi K3. The one thing to know: cheap-per-token is not the same as cheap-per-completed-task if the model needs more turns to land a change. Measure on your own repo, not a leaderboard — the cost-per-fix comparison in Kimi K3 vs Opus vs GPT-5.6 is the honest framing.

Claude Code — the premium you keep for the paths that must land#

Claude Code isn't competing on price and doesn't pretend to. It needs a Max subscription, or Opus 5 on the API at $5/$25 per million tokens (input/output). What you're renting is the model you trust when a change has to be correct — the reliability-critical paths where a wrong diff costs more than a month of any subscription. The move most solo founders land on: run cheap by default, and spend Claude Code's premium only where a failure is expensive. (If Sonnet 5 is your daily model, note its introductory pricing ends August 31 — the money math is in two dated events that raise your AI bill this month.)

Antigravity CLI — no longer the free workhorse#

If you reach for Google's terminal agent out of habit, reset your expectations. Antigravity CLI — a closed-source Go binary invoked as agy, built for async multi-agent runs — replaced Gemini CLI on June 18, but its free quota is about 20 requests a day, down from Gemini CLI's old 1,000. Enterprise accounts on Gemini Code Assist with API-key auth were spared; individual free and Pro/Ultra users were not, and the switch silently changed MCP config in a way that broke CI for people who'd scripted against the old binary. For a founder optimizing for cost, Antigravity is only interesting if you're already committed to Google's stack — the free-workhorse era of Gemini CLI is over. (For the IDE-agent side of Google's shift, see Antigravity vs Cursor vs Claude Code.)

The floor, if you own the hardware#

There's one tier below all of this: open weights on your own GPU. Kimi K3's weights are open, and open coding models like Qwen3-Coder run on a single card. You pay for compute, not per-token rent — and, more to the point this quarter, no vendor can flip a price or pull a login on you the way Google just did. That resilience is the real argument for the open path, not the raw arithmetic.

What to actually do#

The cheapest coding agent is free. The right coding agent is a routing decision — free for the throwaway work, cheap for the volume, premium for the paths that earn it. Wire a swappable client once, and every one of these prices becomes a config change instead of a migration.