OpenAI's Codex and Anthropic's Claude Code have converged on pricing, but they remain fundamentally different tools. Both now charge $10 per million input tokens and $50 per million output tokens for their frontier models, a shift that erases the cost advantage either held just days ago. The convergence happened fast: Anthropic shipped Claude Fable 5.1 on September 1, 2026, and OpenAI followed with GPT-6 Astra on September 3, making it Codex's default on September 4.

For developers and teams evaluating coding agents, this pricing parity means the choice between them now hinges on workflow philosophy rather than budget. "Any Claude Code vs Codex comparison written before September 2026 is describing models that are no longer the default on either side," noted Nafis Amiri, co-founder of CatDoes, a tool that routes coding tasks across model tiers. The two agents have also narrowed their performance gap. On Terminal-Bench 4.0, a widely used coding benchmark, Astra scores roughly 57.7 compared to Fable 5.1's 55.8, a difference of less than 2 points. On DeepSWE v1.1, another industry benchmark, the two are effectively tied at around 74%.

What's the Real Difference Between Claude Code and Codex Now?

The pricing alignment masks a deeper architectural split. Claude Code operates as a supervised, local-first agent that keeps your code on your machine and shows its reasoning at each step. Codex takes the opposite approach: it's built for autonomous cloud execution, where you hand off a task and receive a completed branch or pull request. This distinction shapes everything from how you interact with each tool to how much control you retain over the process.

Claude Code runs across five surfaces that share a single engine and configuration file: a command-line interface, desktop apps for macOS, Windows, and Linux, web and mobile apps, a VS Code extension, and a JetBrains plugin. Your repository never leaves your machine. Codex, by contrast, spans even more surfaces, including a Rust-based CLI, IDE extensions for VS Code, Cursor, Windsurf, JetBrains, and Xcode, plus integrations with GitHub, GitLab, Slack, and Linear. The Codex desktop app was folded into ChatGPT's desktop application in July 2026, consolidating both products under one roof.

How to Choose Between Claude Code and Codex for Your Team

• Supervision and Customization: Claude Code offers deep hook and skill customization, checkpointing before every change, and permission granularity across six modes from fully manual to fully autonomous. If your team needs to review and approve each step, Claude Code's supervised workflow is the better fit.
• Autonomous Execution: Codex is designed to run unattended and fire off parallel tasks in cloud sandboxes. If you want to delegate a task and return to a finished pull request, Codex's autonomous model aligns with that workflow.
• Entry Cost and Free Tier: Codex offers a genuinely free tier and an $8 monthly Go plan, while Claude Code starts at $20 per month for Pro. If budget is a constraint, Codex has a lower barrier to entry.
• Configuration Standards: Codex supports AGENTS.md, a cross-tool configuration standard that lets you define agent behavior per directory. Claude Code does not yet support this standard, despite a feature request drawing roughly 5,000 GitHub reactions.
• Context Window and Mid-Tier Pricing: Both offer roughly 1 million tokens of context at the frontier. Claude Code's mid-tier option, Sonnet 5, costs $2 per million input tokens and $10 per million output tokens with the full 1 million token window, making it the least expensive million-token frontier-class model either lab ships.

What Do the Benchmarks Actually Tell Us?

Performance metrics reveal that the two models are now competitive rather than clearly ranked. Terminal-Bench 4.0, a benchmark that measures coding agent performance, shows Astra at 57.7 versus Fable 5.1 at 55.8. For context, Claude Opus 5, the previous default for Claude Code, scored 52.3 on the same benchmark, meaning the frontier models have both improved. On DeepSWE v1.1, Astra and Fable 5.1 both hover around 74%, effectively tied. Neither Anthropic nor OpenAI publishes SWE-bench Verified scores for their current models anymore, so direct comparison on that metric is no longer possible.

The benchmark convergence reflects a broader trend in the coding agent space: frontier models from major labs are becoming increasingly competitive. Most experienced developers now run both Claude Code and Codex, routing different types of tasks to each based on workflow needs rather than performance alone.

What's Changing in the Coming Weeks?

Anthropic announced that starting September 14, 2026, it will permanently raise standard weekly Claude Code limits by 25% against the pre-May baseline. However, because a temporary 50% promotion has been running since May 13, this change works out to a 17% reduction compared to what subscribers have had all summer. The adjustment reflects Anthropic's effort to balance accessibility with resource constraints as usage grows.

For teams evaluating these tools, the September 2026 updates represent a reset. The frontier models are now price-matched, performance-matched, and available across comparable surfaces. The choice between them depends on whether your team values supervised, customizable workflows with local code storage, or autonomous cloud execution with broader integration support. Most organizations find value in running both, routing work to each based on task type and team preference.