OpenAI Codex CLI vs Anthropic Claude Code: Pick by What You're Doing

OpenAI Codex CLI vs Anthropic Claude Code: Pick by What You're Doing

TL;DR: Last month a friend asked me about tooling for his team. Five engineers — three were Claude evangelists, two were on Team Codex, neither side could close. The sprint slipped two weeks while they argued. I told him: you're not picking between A and B. Pick by task. He paused: "we can do that?" This is the easiest trap to fall into in 2026 AI tooling: treating "Codex vs Claude" as a tribal question. It isn't. The two have clear, measurable differences across specific scenarios — context window, price, API protocol, ecosystem. This post is a real decision tree by task: 7 concrete scenarios with picks, plus a dual-key setup that lets you run both without picking sides. Read it once and the next sprint argument has data behind it.

Model Lineup Snapshot (May 2026)

OpenAI shipped GPT-5.5 (gpt-5.5) on April 23, 2026 — $5/$30 per 1M tokens, 1M context, state-of-the-art on coding benchmarks.

But Codex CLI still defaults to gpt-5.3-codex ($1.75/$14, 400K context) — OpenAI's codex-tuned model that's 3x cheaper. In the scenarios below, "Codex CLI" means this unless noted otherwise.

Switch to gpt-5.5 manually when you need deeper reasoning or a longer context window (>400K tokens) — 3x more expensive but state-of-the-art.

Anthropic's lineup is unchanged: Sonnet 4.6 / Haiku 4.5 / Opus 4.7.

Table of Contents

1. Four-dimension comparison: get the positioning right
2. Seven real scenarios, picked by task
3. The decision tree (copy-paste ready)
4. Dual-key setup: run both in parallel
5. Three traps to avoid
6. FAQ
7. Further reading

Four-dimension comparison: get the positioning right

Get "what's actually different" on the table early. The four dimensions below — price, context, protocol, ecosystem — drive 90% of the selection decision.

Price (as of 2026-05, per 1M tokens)

Quick read: GPT-5.3 Codex's input is roughly 40% cheaper than Sonnet 4.6; output is essentially tied. Haiku 4.5 is the cheapest at this tier. Opus 4.7 is the most expensive — but the only one that handles complex architectural reasoning well.

Context window

Quick read: Want to feed an entire codebase in for a holistic decision? Claude Sonnet 4.6 is in a class of its own (real comparison in scenario #2 below).

API protocol

CodeGateway's gateway proxies both. One `sk-cg-` key works for both. The difference is the client SDK / CLI — Codex CLI is hardcoded to the Responses API, Claude Code to the Messages API. Switch via env vars.

Ecosystem and features

Quick read: solo dev, doesn't care about skill system → Codex is sufficient and cheaper. Multi-person team, engineering culture, wants to share skill packs → Claude Code is smoother.

Seven real scenarios, picked by task

Scenario 1: Refactoring a 30K-LOC mid-sized Python project

• The work: delete dead code, rename, split functions, add type hints, run tests.
• Pick Codex CLI + GPT-5.3 Codex: input is cheaper ($1.75 vs $3), the work is mechanical-heavy, and refactor density doesn't need top-tier reasoning. Sub-agents run 9 tasks in parallel — full evening run is around $1–3. See the Codex playbook for legacy cleanup.
• Counter-example: if the project is closer to 100K LOC and needs "read the whole thing before deciding" → switch to Sonnet 4.6 (1M context, fits in a single call).

Scenario 2: Long-context document analysis (500-page PDF / full PRD set)

• The work: read everything, extract elements, risks, timelines.
• Pick Claude Sonnet 4.6: 1M context dominates — fits the entire document set in a single call without chunking. Codex would have to slice into 5–8 chunks; total tokens may even exceed Claude's single 1M call.
• Prompt cache hits drop input pricing to ~10%, making long-doc Q&A loops very economical.

Scenario 3: Writing tests (spec available)

• The work: backfill 80% coverage on a mid-sized module from spec.
• Pick Codex CLI + GPT-5.3 Codex: writing tests is medium reasoning + heavy output. By price, Codex output ($14) is just under Sonnet's ($15). One evening backfilling 12K test LOC: estimated $1–2.
• Counter-example: spec unclear and you need the AI to draft spec? Sonnet 4.6's long context plus reasoning depth is steadier.

Scenario 4: UI / React component generation

• The work: turn a Figma description or PRD into React components.
• Either Codex CLI + GPT-5.3 Codex or Claude Code + Sonnet 4.6 — basically a tie. Both write components fluently and know component-library styles.
• If your team has a design system (custom tokens, specific Skills) → Claude Code, because Skills like frontend-design commit into the repo for team reuse.
• Pure solo project → cheapest = Codex CLI.

Scenario 5: Cross-file / cross-service field rename

• The work: rename userId → accountId, affecting 6 repos and 30+ files.
• Pick Claude Code's multi-agent: sub-agent decomposition (grep / patch backend / patch frontend / run tests / collect failures) is significantly more mature. The Hooks system can install a "firewall" to auto-block edits to test config.
• Codex can also do this, but the sub-agent maturity is a step behind.

Scenario 6: Architectural decisions (multi-service design / DB migration plan)

• The work: produce a database-migration plan spanning 5 services with rollback.
• Pick Claude Opus 4.7: architectural reasoning is exactly what Opus does in one shot. A single Opus call ($5/$25) looks expensive — but it saves you 3 rounds of Sonnet back-and-forth + rework. Decision-class tasks are 70% thinking, 30% writing — Opus's ROI dominates at that ratio.
• Counter-example: rarely use Opus for daily code tasks. The bill explodes.

Scenario 7: High-concurrency batch tasks (CI lint / full test attribution)

• The work: CI fires, scan 50 PR diffs, post comments.
• Pick Claude Haiku 4.5 or GPT-5.4-mini: unit pricing $1/$5 vs $0.40/$1.60. Mechanical task density; reasoning depth doesn't matter.
• Critical: dedicated CI key (named ci-<repo>), set RPM cap and monthly spend cap. Avoids one stuck loop blowing the budget.

Scenario quick-pick table

The decision tree (copy-paste ready)

Walk top-down by the task's most prominent feature:

What's the dominant feature of your task?

├─ Input exceeds 200K tokens?
│   └─ Claude Sonnet 4.6 (1M context, only one in this league)
│
├─ 100% mechanical / repetitive (lint, format, simple gen)?
│   └─ Haiku 4.5 or gpt-4.1-mini (cheapest unit prices)
│
├─ Deep architectural reasoning (multi-service design, hard debug)?
│   └─ Claude Opus 4.7 (one-shot investment beats N rounds of rework)
│
├─ Cross-N-files refactor / rename / cross-service work?
│   └─ Claude Code multi-agent (sub-agent + hook + skill maturity)
│
├─ Mid-project cleanup / writing tests / UI components?
│   ├─ Cost-sensitive → Codex CLI + gpt-5.3-codex (input ~40% cheaper than Sonnet)
│   └─ Team-engineering-focused → Claude Code + Sonnet 4.6 (skill system in repo)
│
└─ Genuinely unsure?
    └─ Sonnet 4.6 as the default (top balance), drop sub-agents to Haiku 4.5 for cost

Dual-key setup: run both in parallel

The most efficient setup isn't "pick A or B" — it's run both, switch by task. CodeGateway exposes both Anthropic and OpenAI under one account, so dual-running needs no second account.

Step 1: Issue two keys (label by purpose)

Dashboard → API Keys → Create Key, twice:

• claude-laptop (used by Claude Code)
• codex-laptop (used by Codex CLI)

Technically a single key works for both — but separate naming has operational benefits: per-key Logs filtering, per-key alerting, per-key rotation on leak.

Step 2: Env vars per "active tool"

# File 1: ~/.claude-env
export ANTHROPIC_BASE_URL="https://api.codegateway.dev"
export ANTHROPIC_API_KEY="sk-cg-claude-key-xxx"
export ANTHROPIC_MODEL="claude-sonnet-4-6"

# File 2: ~/.codex-env
export OPENAI_BASE_URL="https://api.codegateway.dev/v1"
export OPENAI_API_KEY="sk-cg-codex-key-xxx"

Run Claude Code → source ~/.claude-env. Switch to Codex → source ~/.codex-env.

Step 3: Project-level overrides (recommended)

Different projects, different default tools. .envrc (with direnv) + git-ignored:

# legacy-cleanup-project/.envrc
source ~/.codex-env
echo "✓ Codex environment loaded for legacy cleanup"

# new-saas-project/.envrc
source ~/.claude-env
echo "✓ Claude environment loaded for SaaS work"

cd into a project, env auto-loads. Less manual env switching, fewer "wrong env left over from last shell" bugs.

Step 4: Per-key billing visibility

CodeGateway Dashboard → Logs → filter by key. You see:

• claude-laptop 30-day token usage
• codex-laptop 30-day token usage
• Which key spent the most on which model

Three traps to avoid

Trap 1: "Claude is smarter than GPT" / "GPT is faster than Claude"

Wrong. Overall capability differences are < 10% on public benchmarks. Single-scenario differences are significant (e.g., long context). Comparing overall is market narrative; for your specific project, it's irrelevant.

Right framing: for the specific task you have this week, which model's strength zone covers it?

Trap 2: "Cheaper is always better"

Wrong. Per-token, gpt-5.3-codex is 40% cheaper than Sonnet 4.6. But if Codex needs 3 rounds to get to a result Claude gives you in one round, end-to-end cost flips at 1.5x.

Right framing: cost per completed task, not per token.

Trap 3: "My team standardizes on X"

Wrong. The ops gain from standardization is much smaller than the hidden cost of "tool-task mismatch." Developers themselves are heterogeneous — backend optimizing SQL with Sonnet vs frontend generating components with Codex; no reason to force them to use the same.

Right framing: standardize on infrastructure (one CodeGateway key, one billing line, one monitoring dashboard); leave tool and model choice to the individual / project.

FAQ

Q: Codex CLI and Claude Code on the same key — really?

A: Yes. CodeGateway routes by request endpoint at the gateway layer — /v1/responses to OpenAI, /v1/messages to Anthropic. Same key works both ways.

Q: Then why bother with two named keys?

A: Not technically necessary — operationally necessary. Per-key Logs grouping, per-key spend alerts, per-key rotation on leak. One mixed key obscures all those dimensions.

Q: How does CodeGateway's tier markup work with Codex + Claude mixed? Cumulative. Token usage across upstream models all rolls into your 90-day spend window. New accounts start at 1.5x; cumulative $10 drops to 1.4x; floor at 1.2x. Mixing actually hits lower tiers faster than going single-sided. See Tier markup explainer.

Q: Can I run Codex and Claude Code at the same time?

A: Yes. Two shells, each with its own env. RPM limits are per-key (not per-account), so two keys = double RPM headroom.

Where does Cursor fit? It works with Claude and OpenAI. Cursor is an IDE — different shape from CLI agents. Cursor wins at in-editor interactive editing, inline completion, UI feedback; Codex CLI / Claude Code wins at long-horizon automation. Running all three is sensible — Cursor for short edits, CLI for big tasks. CodeGateway keys work in Cursor too (Settings → set base URL + API key).

Q: Can I switch later?

A: Yes, and cheaply. CodeGateway's one-key-dual-protocol means swapping tools doesn't require new API keys or data migration. This is the core CodeGateway advantage — avoid single-vendor lock-in.

Q: Is Claude's 1M context really useful?

A: Depends on the task. Daily 5K–30K-token tasks don't need it. But when a task must fit a whole codebase / full PRD / complete incident log into one inference, 1M is 3.7× GPT-5.3 Codex's 272K — directly determines whether the task is feasible. Concrete threshold: single inputs > 200K tokens → strongly prefer Sonnet 4.6.

Architectural decisions really worth Opus? $5/$25 looks pricey. Pricey per call, possibly cheaper end-to-end. A DB migration decision via Sonnet may take 3 rounds ($1–2 each) plus one rework ($3–5), totaling $7–12; Opus one-shot costs $3–5. Don't use Opus for daily code, but for decision-class tasks the ROI is high.

Further reading

• Cleaning up legacy code with Codex CLI — Codex hands-on, pairs with scenario 1 above
• The complete Claude Code configuration guide — Claude engineering setup
• Advanced Claude Code: sub-agents, MCP, cost optimization — multi-agent + MCP + cost
• Claude Code connection timeout troubleshooting — long-task drops, applies to both tools
• An honest receipt: 16 blog hero images for $0.92 — same useful-problem + receipt approach for image gen
• Top-up and billing guide
• Tier markup explainer — 90-day rolling, floor 1.2x
• Anthropic — Claude pricing
• OpenAI — Pricing

Picking a tool isn't picking a tribe. Both are good — what matters is getting specific about the task: context length, reasoning depth, repetitiveness, team-engineering needs. These features tell you which tool's strengths line up with you. Next time the team argues, put the task on the table early, then read this tree.

Gemini CLI vs Claude Code: The Third Option

If you thought picking between Codex CLI and Claude Code was already hard enough, 2026 handed you a third player: Gemini CLI.

What it is: Google DeepMind open-sourced Gemini CLI in June 2025. It's a terminal-based AI agent backed by Gemini 2.5 Pro — not a VS Code plugin, not a browser chat interface, a proper CLI agent that competes directly with Claude Code and Codex CLI in form factor and use cases.

So the real question is no longer "Codex vs Claude" — it's "Codex vs Claude vs Gemini, and when."

What Gemini CLI Gets Right

1. The lowest entry barrier of the three

Sign in with a personal Google account and you're running. The Gemini API free tier includes 15 requests per minute — enough for personal projects, light exploration, and evaluation runs, at zero cost. Neither Claude Code nor Codex CLI offers a free tier. If you want to trial a terminal AI agent without committing a credit card, Gemini CLI wins by default.

2. Native Google Search integration

Gemini CLI can call Google Search mid-task. When your agent needs to look up updated API docs, chase a library's latest release notes, or verify a deprecation, it does that inline without MCP setup or external hooks. Claude Code and Codex CLI both need MCP or custom hook plumbing to get equivalent behavior. For research-heavy tasks or working with fast-moving ecosystems, this is a real edge.

3. 1M token context window

Gemini 2.5 Pro ships with a 1M token context window — on par with Claude Sonnet 4.6, and 3.7× GPT-5.3 Codex's 272K cap. Feeding an entire codebase in one shot? Gemini CLI can do it. This puts it firmly in "whole-repo analysis" territory that Codex CLI can't match without chunking.

4. Native multimodality

Text, images, screenshots, video frames — Gemini 2.5 Pro handles all of it natively. For front-end tasks where you're feeding in Figma mockups or UI screenshots alongside code, this is a friction-free workflow that Claude Code and Codex CLI don't fully replicate.

Where Gemini CLI Falls Short

1. Engineering depth is noticeably behind

Claude Code has a mature Skills system, a three-layer Hooks architecture (PreToolUse / PostToolUse / Stop), and first-class MCP support baked in. Codex CLI has a reasonable sub-agent system. Gemini CLI is several iterations behind both on workflow orchestration, multi-agent scheduling, and team-level reuse patterns. As of mid-2026, it reads more like "powerful personal tool" than "team engineering infrastructure."

2. Code-editing precision trails Claude

For cross-file refactors, cross-service field renames, and complex multi-repo operations, Claude Sonnet 4.6 still delivers tighter, more controlled edits. Gemini 2.5 Pro has a documented tendency to modify code outside the stated scope — engineering teams report more cleanup noise. When control matters, Claude Code remains more predictable.

3. Deep reasoning: Opus still has the edge

For decision-class tasks — multi-service migration planning, architectural tradeoffs, anything where 70% of the work is "thinking and not writing" — Claude Opus 4.7 still wins. Gemini 2.5 Pro is benchmark-competitive with Sonnet 4.6, but in real-world architectural sessions, engineers report Opus delivers more "one-shot-complete" results before rework.

Three-Way Quick-Pick Table

| Dimension | Claude Code | Codex CLI | Gemini CLI |

When to Actually Pick Gemini CLI

Pick it when:

• You're evaluating a terminal AI agent for the first time and don't want to put a card down — free tier makes the math easy

• Your task requires real-time web search or live documentation lookup inline with code work — no MCP config, it just works

• You're doing front-end or design-adjacent tasks with heavy screenshot / mockup input — multimodality is friction-free

Stick with Claude Code when:

• Your team has invested in shared skills, hooks, and MCP integrations → Claude Code's engineering depth compounds over time

• You need architectural decisions made well in one shot → Claude Opus 4.7 is still the most reliable here

• Predictable, scope-controlled edits are critical → Claude's tighter edit behavior wins

Stick with Codex CLI when:

• You want the cheapest input token price for mechanical refactoring and test generation → gpt-5.3-codex at $1.75/M input is still one of the lowest unit prices in this tier

• You're a solo developer, don't care about skills systems, and just want something fast and cheap → Codex is sufficient

The "pick any of the three" architecture

There's a reason this is relevant for CodeGateway users: these three tools run on three separate APIs — Anthropic Messages API, OpenAI Responses API, and Google Gemini API. Right now, one CodeGateway key covers Claude's full lineup and OpenAI's full lineup under one account, one billing line, one log dashboard. Gemini integration is on the roadmap.

The goal: one key, any model, any CLI tool. You pick by task, not by who manages your API subscription. No re-issuing credentials every time a new model ships, no juggling three separate dashboards for spend monitoring. That's the actual value of a gateway layer in a world where the "right tool" is changing every quarter.