Short version: three terminal coding agents now start at roughly the same price — Kimi Code's new $19 tier, Claude Code's $20 Pro plan, and the free Codex CLI — so the decision isn't the sticker. It's what you're renting. Codex CLI is free and folded into a ChatGPT plan you may already pay for. Claude Code buys the strongest default coding model (Sonnet 5). And Kimi Code — which switched on paid USD tiers this week, right as Moonshot's 2.8-trillion-parameter K3 weights land — is the cheap challenger with a real escape hatch: an open model you can eventually run yourself. There's one catch, and it isn't price.
The three agents, on one card#
| Kimi Code | Claude Code | Codex CLI | |
|---|---|---|---|
| Entry price | Adagio free / Moderato $19 | Pro $20 (Max $100/$200) | Free, in any ChatGPT plan |
| Default model | K-series flagship (K3 rolling out) | Claude Sonnet 5 | GPT-5.6 |
| BYO-key path | Anthropic-compatible, kimi-k3 $3/$15 | Anthropic API | OpenAI API, gpt-5.6-terra $2.50/$15 |
| Open weights | Yes — K3 full drop July 27 | No | No |
| Capacity risk | High — new subs paused July 18 | Low | Low |
All three are the same shape of tool: an agent that lives in your terminal, reads and writes files, runs commands, plans across steps, and speaks MCP. The differences are cost model, model quality, and — the part most comparisons skip — whether you can get a reliable seat.
What you're actually renting#
Codex CLI is the quiet value play. The CLI is free, open-source, and included with every ChatGPT plan — Free, Go ($8), Plus ($20), Pro ($100/$200). If your team already pays OpenAI, Codex adds zero to the bill; it draws from the same rolling 5-hour usage window as Codex web and IDE, so a heavy terminal session competes with your other Codex work. Point it at your own OpenAI key instead and you pay per token (gpt-5.6-terra at $2.50/$15 per 1M), with no window cap.
Claude Code is the quality default. It ships with a Claude plan — Pro $20/mo, Max $100 or $200/mo — and runs Sonnet 5 out of the box, which remains the strongest default coding model of the three. If you want the least-fiddly setup and the best odds on a hard refactor landing first try, this is the safe pick.
Kimi Code is the challenger, and it moved this week. Moonshot turned on official USD subscriptions on July 24 — Adagio (free, tight limits), Moderato $19, Allegretto $39, Allegro $99, Vivace $199 — each step buying more usage inside rolling 5-hour windows (roughly 300–1,200 calls per window, up to 30 concurrent). The free tier runs an older model, not the flagship, so real work means a paid seat or the API.
The escape hatch only Kimi has#
Here's the strategic difference, and it's the reason to care about Kimi Code even if you never open its CLI:
Kimi Code's backend speaks an Anthropic-compatible API, and its weights are open. You can drive K3 from inside Claude Code today, and self-host it once the weights drop July 27.
Concretely, you can point Claude Code straight at K3 — set ANTHROPIC_BASE_URL to https://api.moonshot.ai/anthropic, ANTHROPIC_AUTH_TOKEN to your Moonshot key, and ANTHROPIC_MODEL to kimi-k3. (One trap: a leftover ANTHROPIC_API_KEY silently overrides ANTHROPIC_AUTH_TOKEN and breaks it; the /model menu won't show Kimi, so confirm with /status.) That makes K3 a drop-in you can A/B against Sonnet 5 in the same tool. On the API, K3 bills at $3/$15 per 1M — about 40% under Opus 4.8 — and once the full 2.8T open weights land Sunday, it's the only one of the three you can eventually run on your own metal. Neither Claude Code nor Codex gives you a weight to move.
The catch that isn't price#
Kimi Code's problem right now is capacity, not cost. On July 18, Moonshot said K3 demand had pushed "close to the limits of our current capacity" and paused new subscriptions to protect existing users, reopening spots in batches as it adds GPUs. So the risk isn't the model or the $19 — it's whether you can get, and keep, a reliable slot during the launch crush. That's the backdrop to why frontier labs are locking up multi-year compute leases: serving a 2.8T model at scale is a capacity problem before it's a pricing one. For anything you can't afford to have rate-limited mid-task, keep a fallback agent configured.
The decision, in one line#
- Codex CLI if you already pay for ChatGPT — it's free and included, and the cheapest capable agent by default.
- Claude Code if coding quality and low-friction setup matter most — Sonnet 5 is the strongest default, and Max ($100/$200) buys the headroom for all-day agent runs.
- Kimi Code if you want the cheapest capable agent and the escape hatch — an Anthropic-compatible API you can drive from other tools plus open weights you can self-host — and you can secure a slot.
And do the one thing that survives every price and capacity swing this summer: keep your agent workflow portable. Because all three speak MCP and Anthropic- or OpenAI-compatible APIs, the switching cost between them is mostly config, not rewrite — which is exactly the leverage you want when the cheapest capable model changes twice in a week. Pick a default, wire a fallback, and measure them on cost per completed task, not sticker price. If you're weighing the subscription math more broadly, we broke it down in which AI coding subscription a solo founder should buy.
