DEV Community

Evan-dong
Evan-dong

Posted on • Originally published at evolink.ai

Codex vs Claude Code vs Cursor: What Changed in May 2026 and What Can Be Routed

Cursor cloud agent development environments

A common setup now is Codex CLI for scaffolding, Claude Code for refactoring, and Cursor for IDE or cloud-agent workflows. This month all three shipped infra updates — on different dates, solving different problems.

Here is what changed, what can be wired through one endpoint, and what has to remain separate.

The Updates

Codex CLI — Windows Sandbox (May 13)

Codex now has real OS-level sandboxing on Windows. Dedicated user accounts (CodexSandboxOffline/CodexSandboxOnline), per-account firewall rules, helper binaries for privilege boundaries. Linux already had seccomp/bubblewrap; Windows finally caught up.

Previously, Windows sandbox attempts used synthetic SIDs and proxy-based network blocking — programs could bypass them by implementing their own networking stack.

Claude Code — Doubled Five-Hour Rate Limits (May 6)

Anthropic doubled Claude Code's five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans. Also removed the peak hours limit reduction for Pro and Max.

This was announced May 6, a week before the other updates. If you were splitting refactoring tasks to stay under the five-hour cap, you have roughly 2x headroom.

Cursor — Cloud Agent Environments (May 13)

Cloud environments with multi-repo support, Dockerfile config with build secrets, layer caching (70% faster on cache hits), version history, rollback, scoped egress/secrets, audit logging.

Dockerfile auto-config is in private beta, rolling out to Enterprise teams.

Quick Comparison

Codex Claude Code Cursor
Updated May 13 May 6 May 13
What Windows sandbox 2x five-hour limits Cloud dev environments
Runs Local sandbox Local terminal Cloud + IDE
Multi-repo No No Yes
Isolation OS-level None Cloud containers
Best for Scaffolding, Windows Refactoring, terminal End-to-end delivery

The Problem When You Use More Than One

Multiple provider keys, billing dashboards, and rate-limit surfaces. When Claude Code hits its five-hour cap, no automatic fallback happens by default. When you want to compare spend across routable CLIs, you are exporting data from separate consoles.

One Gateway Setup

Point the routable CLIs at one endpoint:

# Codex: ~/.codex/config.toml
[api]
base_url = "https://api.evolink.ai/v1"
Enter fullscreen mode Exit fullscreen mode
# Claude Code
export ANTHROPIC_BASE_URL="https://api.evolink.ai"
export ANTHROPIC_AUTH_TOKEN="your-evolink-key"
Enter fullscreen mode Exit fullscreen mode
# Gemini CLI: ~/.gemini/.env
GEMINI_API_KEY=your-evolink-key
GOOGLE_GEMINI_BASE_URL=https://api.evolink.ai/
Enter fullscreen mode Exit fullscreen mode

What you get: unified cost dashboard, automatic fallback on 429/5xx, model switching via routing rules.

Note: Cursor cloud environments use their own backend — not routable through third-party gateways yet. Local Cursor IDE completions can be routed.

Full walkthrough: One Gateway for 3 Coding CLIs

Watch Out For

  • Config precedence: env vars override config files. If you set ANTHROPIC_BASE_URL globally and have a project .claude/settings.json, behavior depends on tool version.
  • Auth errors: 401 from gateway usually means old provider key is still being sent. Restart terminal after env var changes.
  • Upstream limits still apply: gateway adds fallback, but Anthropic five-hour cap and OpenAI daily quotas are still upstream.

Sources

Top comments (0)