Claude Code vs Codex CLI: Which Terminal Agent in 2026?

Claude Code and Codex CLI are the two terminal coding agents most engineers actually shortlist in 2026. From a distance they look alike — both live in your terminal, read your repo, edit files, and run commands — but they speak different protocols, reach different model families, and reward different working styles. This comparison is about daily engineering work, and it ends in a pick-by-workflow verdict rather than a crowned winner. It also covers the part most comparisons skip: running both through one Router One account, so trying them side by side doesn't require holding two subscriptions.

One scoping note up front: both of these are terminal agents, not editors. If your real question is "agent in the terminal vs AI in the IDE," read Cursor vs Claude Code first. This post assumes you've already decided you want the terminal. If a third terminal agent belongs on your shortlist, OpenCode vs Claude Code vs Codex CLI adds OpenCode to this same comparison. For DeepSeek's own agent, DeepSeek Harness vs Claude Code vs Codex CLI puts DeepSeek Harness next to both.

Two agents, two philosophies

Claude Code is Anthropic's terminal agent. You run claude in a repo, describe a task, and it explores the codebase on its own — reading files, editing them, running tests, fixing failures, iterating. Its default posture is "do the work, then show you the diff": it asks at tool-use boundaries, but it is built for long autonomous stretches that touch fifteen files before coming back to you.

Codex CLI is OpenAI's terminal agent. Two things define it. First, it is native to the Responses API wire format — the newer protocol OpenAI built for agentic tool use — rather than classic Chat Completions, which is why many generic relays fail with it while a gateway that implements wire_api = "responses" works — see Codex and the Responses API. Second, its execution model is sandbox-first: commands run inside a sandbox with explicit approval modes, and the agent has to ask before escalating. The leash is shorter by design.

That difference in default posture is the honest one-line summary: Claude Code optimizes for autonomy, Codex CLI optimizes for containment. Which one is "better" depends entirely on whether the task in front of you deserves a delegate or a supervised executor.

What each tool is actually good at

Claude Code's strengths:

• Deep multi-file agentic edits. It holds a plan across a whole module — refactors that touch 15–30 files, iterate-until-tests-pass loops, migrations that need the same idea applied consistently everywhere.
• The hooks and skills ecosystem. Hooks fire shell commands at lifecycle events (before/after tool use), skills package repo-specific workflows into reusable commands, and CLAUDE.md gives every session standing project context. Teams that invest here get an agent that behaves like it has been onboarded.
• Terminal-native, editor-agnostic. It operates at the filesystem level, so it pairs with Vim, VS Code, or anything else; you review with git diff.

Codex CLI's strengths:

• Responses-API native. Structured tool calls and reasoning summaries come through the protocol as first-class objects, not as prompt conventions layered on Chat Completions.
• Sandboxed execution style. The approval-mode model makes it comfortable to run on codebases where an unsupervised rm or an accidental push would be expensive. It is the easier tool to hand to someone cautious.
• The GPT family on one config. GPT-5.6 Sol is the strong generalist; GPT-5.6 Terra is listed at a lower rate for small, well-scoped edits, and GPT-6 Astra sits at the top of the range — switching is one model line.

The model boundary, stated honestly

Here is the part vendors blur and we won't. Through Router One, one account, one wallet, and one gateway key serve both tools — but each tool only reaches the models that speak its own protocol.

Claude Code uses the Anthropic Messages endpoint and currently listed Claude models. Codex CLI uses native Responses and currently listed GPT models, including GPT-6 Astra and the GPT-5.6 family. DeepSeek V4 Pro left the catalog on 2026-09-05; deepseek-v4.1-flash and deepseek-v4-flash have been listed since 2026-09-10 and answer on /v1/chat/completions, /v1/messages and /v1/responses, so either CLI can reach them. Grok chat ids answer on /v1/responses, so they work in Codex CLI; Claude ids are rejected on /v1/responses. Check each model’s endpoint list before configuring either CLI.

What is shared is everything around the calls: the same wallet funds both, the same dashboard shows both, and per-request traces record what each tool spent. That's the practical meaning of "one gateway for both."

Side by side

Setting up both through Router One

Both configurations take about two minutes, and the one-click install commands do them for you. Manually, Claude Code is an environment pair — the base URL plus your key sent as the bearer token:

export ANTHROPIC_BASE_URL=https://api.router.one
export ANTHROPIC_AUTH_TOKEN=sk-your-api-key

Codex CLI is a custom provider in ~/.codex/config.toml using the env_key mechanism — the config names an environment variable, and Codex reads your Router One key from it:

model = "gpt-5.6-sol"
model_provider = "router"

[model_providers.router]
name = "Router One"
base_url = "https://api.router.one/v1"
wire_api = "responses"
env_key = "ROUTER_ONE_API_KEY"

Then export ROUTER_ONE_API_KEY=sk-your-api-key in your shell profile. The model line matches what the one-click install writes: gpt-5.6-sol, the bare name from Codex's own model list, which sits in the Premium models tier of the Pro, Max and Ultra plans as of the 2026-10-01 plan response. The CLI setup guide has the complete version of both, including Windows. OpenCode, the third terminal agent, points at the same key — the OpenCode setup guide has its config.

Cost: one wallet instead of two subscriptions

Run natively, these tools sit on two separate vendor relationships — two billing pages, two payment methods, two places to check what you spent. Through Router One both become pay-per-token against one prepaid wallet, which changes three practical things:

• Trying both is cheap. Create one key per tool, set a maxSpend hard cap on each key's wallet charges — say $10 — and run a real week of work through each agent. When a key hits its cap, its wallet-billed requests stop, so it cannot charge your wallet past that amount and your other key keeps working.
• Check each request's cost. Dashboard → Logs records the model, request ID, tokens, cost, latency, and status. Save request IDs alongside the tool name in your client or application logs to reconcile workload costs. The cost tracking page describes the available fields.
• Rates are the posted per-token line. Pricing starts as low as 10% of official provider list prices (up to 90% off) on select models — live per-model rates are on the models page. If you prefer a flat monthly number, Pro/Max/Ultra subscriptions exist alongside the wallet (pricing). As of the 2026-10-01 plan response no plan tier lists Claude Opus 5.5, which Claude Code 2.1.280 and later send when no model is set, or Claude Sonnet 5.5, which the sonnet alias sends from Claude Code 2.1.284, so those calls bill per token to the wallet. Plan holders set ANTHROPIC_MODEL=claude-opus-5, ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-5 and ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-5 to keep Claude Code on plan quota; Claude Code in China has the full setup.

One reliability note that applies to both tools: when an eligible upstream error occurs — a 5xx, timeout, 429, or 529 — Router One can retry the request on another healthy provider serving the same requested model. Not every error is retryable and this is not a zero-downtime guarantee; long agent sessions should still tolerate the occasional failed call and use bounded exponential backoff on retries.

For the Codex-specific version of this argument — why a coding CLI doesn't need a consumer chat subscription at all — see Codex CLI without ChatGPT Plus.

Using both from mainland China

Both tools call vendor endpoints that are unreliable from Chinese ISPs without a VPN, and both vendors bill in ways that assume a foreign card. Router One's endpoints are directly reachable from mainland China with no VPN, and the wallet tops up by card or Alipay through one hosted checkout, or USDT/USDC on six chains (Tron, BSC, Ethereum, Polygon, Base, Arbitrum) — no US credit card required. The dedicated setup pages are Claude Code in China and Codex CLI in China.

FAQ

Can Claude Code use GPT models through Router One, or Codex CLI use Claude models? No. Claude Code uses the Messages endpoint for listed Claude models; Codex CLI uses Responses for listed GPT models. The account, wallet, keys, and per-request traces are shared. A shared gateway does not make a Claude model callable through Responses or a GPT model callable through Messages.

Can I run Claude Code and Codex CLI on the same Router One account? Yes. One wallet can fund both tools. Create a separate API key for each tool and set its own maxSpend cap; save request IDs with tool names when you need workload cost attribution.

Do I need a Claude subscription or ChatGPT Plus to use these CLIs? No. Through Router One both tools bill per token from a prepaid wallet at posted per-model rates, with no consumer subscription on either side. Router One's own Pro/Max/Ultra plans exist as an optional flat-rate alternative for heavy users.

How do I see which tool spent what? Dashboard → Logs records the model, request ID, tokens, cost, latency, and status. Match those request IDs with tool names saved in your client or application logs to attribute costs.

Do both tools work from mainland China? Yes, when pointed at Router One: the endpoints are directly reachable from the mainland without a VPN, and the wallet tops up via card or Alipay through a hosted checkout, or stablecoins on six chains. The China-specific setup guides are at /claude-code-china and /codex-china.

Verdict: pick by workflow, not by brand

There is no winner to crown, and anyone crowning one is reviewing their own habits. The honest mapping:

• Default to Claude Code if your work is dominated by multi-file changes in an established codebase, you want run-tests-fix-rerun loops handled without supervision, or your team is willing to invest in hooks, skills, and CLAUDE.md to encode its conventions.
• Default to Codex CLI if you value the sandbox-first execution model, your tasks are mostly contained and well-scoped, or the GPT family — GPT-5.6 Sol for general work, the lower-rate GPT-5.6 Terra for small edits — fits your rhythm.
• Run both if you can. They don't conflict, and through one Router One wallet with a capped key each, the experiment costs exactly what the traces say it cost.

Set up a key at router.one, install either tool with the one-click setup, and let two weeks of real work — not a benchmark table — make the call.