The vibe shift on the timeline is turning against Claude Code

OpenAI's Codex is eating Claude Code's lunch*, and the developer community is taking notice.

*Well, maybe not the whole lunch, but definitely taking a bite!

Just six weeks ago, Claude Code was the undisputed king of AI coding tools. Developers couldn't stop raving about its ability to understand entire codebases and handle complex, multi-file edits.

But something shifted over the summer. Now we're seeing a different story unfold on X, where developers are publicly canceling Claude Code subscriptions and switching to Codex en masse.

‍What’s happened:

It looks like Codex might be gaining on Claude Code. We're seeing evidence of this everywhere on X:

• Data‑science author Chris Albon canceled Claude Code to try Codex for a few months.
• Developer Ian Nuttall posted side‑by‑side tests of Codex vs. Claude Code, concluding he "will be looking to use Codex more."
• And product builder Ian Kar switched last week and says Codex is “really good.” Anecdotal? Yes. But the timeline vibe is clear.

Some more examples:

What Changed for Codex

Zoom in. OpenAI rolled out an IDE extension and a stronger CLI, so Codex now runs in your terminal and IDE without extra setup. A recent write‑up details a big update with GitHub workflows (e.g., @codex code review) and broader plan support. Claude Code still shines: it has subagents for task routing, deterministic hooks, and deep MCP integrations. But CC also recently rolled out new monthly limits that just kicked into effect as of the end of August, which could also be cooling CC's hot streak.

Zoom out. It seems that GPT‑5 raised expectations for reasoning‑heavy coding, and Codex met devs where they work (terminal, VS Code, Cursor). Earlier in the summer, some tests even said Claude Code “crushed” alternatives. Sentiment whiplash is normal in this space, but a growing tidal wave is shifting in OpenAI's favor.

So who's actually better? Claude Code or Codex?

If you want to know which one is REALLY winning, check out this great livestream from Ray Fernando, one of our favorite AI coding creators, who debated which tool is best (and also debated Cursor) alongside his guest in this week's episode, Robin Ebers. Both Ray and Robin spend $1K a month on coding tools, so they'll know exactly which tools have the best bang for their buck: 

The High-End / "Money is No Object" Setup: Cursor

This is the top-tier, no-compromise recommendation, referred to as the "Lambo of AI tools."

• Argument for Cursor: Its primary strength is being a complete, integrated ecosystem. It saves you the headache of installing and managing multiple different tools, extensions, and billing plans. You get all models and features in one place. (6:37)

• Actionable Tip: If you are a "baller" and cost is not your main concern, choose Cursor. It gives you access to all the best models (including all variants of GPT-5 and the expensive Claude Opus) in one interface. (10:35)

• Key Feature: Cursor's context management is superior due to semantic indexing. The AI can ask the IDE natural language questions about your codebase (e.g., "Where is the Stripe integration?") to find relevant code, which is far more efficient than just stuffing files into a context window. (19:38)

• Value Proposition: Paid features like BugBot for automated code review are worth the cost because they are simply better and more integrated than the free alternatives available in other tools. (7:59)
• Dream Setup: The ideal setup would be a truly unlimited Cursor Ultra plan. This would remove "token anxiety" and allow you to freely switch to expensive models like Claude Opus when you need a "second pair of eyes" on a problem, without worrying about the cost. (1:11:00)

The Best Value Setup: Codex (via ChatGPT Plus)

This is the recommendation for the best performance and features for the lowest price.

• Actionable Tip: The single best value in AI coding right now is the standard $20/month ChatGPT Plus subscription. This gives you access to the powerful Codex model via its free and "fantastic" native extension for VS Code. (41:30)

• Argument for Codex: Unlike Claude Code (which runs in a simple terminal), the Codex extension offers a deep, native IDE integration. For example, it can see your //TODO comments in your code and offer to automatically implement them with a single click. (41:53)

• Secret Tip: The Codex CLI contains a secret, fourth variation of GPT-5 called "Minimal". It's the fastest version available anywhere and is perfect for blasting through simple implementation tasks after you've already done the high-level planning. (1:15:27)

The Budget-Friendly Premium Setup: Windsurf

This is for users who want a high-quality experience but can't afford the top-tier plans.

• Actionable Tip: If you're on a budget but still want a great tool, the unconditional recommendation is Windsurf. For about $15/month, you can get around 1,000 GPT-5 requests, which is described as "extremely great value." (29:57)

Actionable Workflows & Pro Tips (How to Use the Tools)

These are specific, actionable strategies for combining and using the tools effectively.

• The "Batman and Robin" Strategy (Cost Management): Use your most expensive, powerful tool (like Cursor Ultra) as "Batman" for the most critical development tasks. For less crucial or high-volume tasks like writing documentation, running security checks, or simple refactoring, use your cheaper subscription (like Codex) as "Robin" to avoid burning through your expensive credits. (45:47)
• The Hybrid Planning Workflow (Best of Both Worlds):
  1. Start in Claude Opus (in Cursor or Claude Code) to create a detailed, high-quality, step-by-step project plan. Opus excels at high-level reasoning.
  2. Take that generated plan and execute it step-by-step using the faster and much cheaper GPT-5 Medium in either Cursor or Codex. (17:33)
• How to Use GPT-5 Effectively: Don't get discouraged by its initial slowness. It's building a deep understanding of your code.
  • For complex new features or bugs, use GPT-5 High for the initial prompt to leverage its superior reasoning.
  • For all subsequent steps in the same chat, switch to the standard GPT-5, as it will be much faster now that it has the context. (13:52)
• For Fast, Targeted Fixes: If you are a technical user and have already pinpointed where a bug is, use GPT-5 Low in Cursor. It is extremely steerable and will implement the precise change you ask for with incredible speed. (14:29)

Fernando's Final Verdict: The Tool Tiers

Here is the final summary of which tool to choose based on your needs and budget.

• The "Lambo" (No Compromises): Cursor. (1:25:48)
• The "BMW" (Solid Premium Value): Windsurf. (1:25:59)
• The Best Overall Value: Codex, via the ChatGPT Plus subscription and its native VS Code extension. (1:26:18)

As for Claude Code? It doesn't land cleanly on the final "tool tiers" list because both Ray and his guest, Robin, say that it has shifted from being a top contender to more of a specialist tool for one specific task: high-level planning.

Here is a breakdown of where it fits and the specific arguments made about it:

Claude Code: The Specialist Tool for High-Level Planning

While no longer recommended as a primary, all-day coding assistant, Claude Code is still valued for the powerful reasoning of its top model, Opus.

• Primary Use Case: The speakers agree that Claude Opus is a "beautiful planner." The best way to use Claude Code now is to generate a detailed, high-quality, step-by-step project plan at the beginning of a task. (17:33)
• Actionable Workflow: Use Claude Code with Opus to think through the architecture and create a plan, then switch to a faster and cheaper tool like Cursor with GPT-5 or the Codex extension for the actual implementation. (42:30)

Arguments for Why It's No Longer a Top-Tier Recommendation

The speakers outlined several key reasons why Claude Code has fallen out of favor as a primary tool:

1. It's Not a True IDE Integration: A major criticism is that it's not a native IDE. As Robin states, "It doesn't run in an IDE. It runs in a terminal inside of an IDE," which makes it less powerful than the deep integrations of Cursor or the native Codex extension. (42:18)
2. Recent Performance Issues: Both speakers feel the tool has been "nerfed" recently. They mention it is slowing down significantly and reference Anthropic's public admission that a deployment issue was causing the model to produce "dumber results." (22:34)
3. Prohibitive Cost of Opus: Using its best model, Opus, is extremely expensive. Robin notes that relying on it for planning caused him to hit his five-hour rate limits, even on a paid plan, making it unsustainable for frequent use. (48:35)

1. Lack of an Ecosystem: Compared to its competitors, it lacks a rich ecosystem. For example, they mention that Codex has a mobile app that works seamlessly, whereas trying to use Claude Code on a phone requires a clunky SSH setup. (43:44)

In summary, Claude Code is now seen as a powerful but situational tool. It's the go-to for its exceptional planning capabilities with Opus, but its interface limitations, recent performance drops, and high cost prevent it from being a top-tier daily driver like Cursor or the best-value option like Codex.

Here's why this matters: Fewer tools means less setup tax and lower cost. If one agent spans terminal + IDE + GitHub, teams move faster with fewer context hops.

But there's a counter‑narrative brewing against all agentic coding, too: where’s the shovelware?

Mike Judge is asking the right question atm: if AI coding is a huge productivity boost, where’s the shovelware? We’re not seeing a tidal wave of throwaway apps. Part of this is the shipping problem—it’s easy to “vibe code,” harder to “vibe ship.” Part is last‑mile friction: env setup, secrets, tests, CI/CD, permissions, reviews. Part is reliability: assistants still miss edge cases that break prod. And part is incentives—playing with the tools is fun; hardening code isn’t.

How Claude could win back users (presuming its lost any): Don’t fight Codex only on coding. It's time to lean into being a general agent—an operating‑system layer for work. Nikunj Kothari argues we’re entering the DOS era, where an agent orchestrates your machine and apps through a terminal substrate. Claude already has the pieces (MCP, hooks, subagents). If Anthropic makes wiring tools, policies, and end‑to‑end flows dead simple, the narrative can and likely will flip back again.

So sure, today’s timeline favors Codex. But if Claude Code becomes the cleanest general agent for real‑world workflows, it can very well win the next news cycle.

Today's X timeline favors Codex, and the momentum feels real. After all, these days, momentum is the moat. OpenAI delivered exactly what developers wanted: seamless integration with their existing workflows and transparent pricing through ChatGPT Plus.

But this is still early innings. The AI coding space moves fast, and Claude Code has fought back from worse positions before.

The real test isn't which tool codes better—they're both capable. It's which one becomes essential to how developers actually work. Codex is winning on convenience and integration. Claude Code's comeback will depend on whether it can evolve beyond coding into something broader and more valuable... and perhaps, in the process, grow its user-base exponentially (so beyond developers). Perhaps this will require a "Nano Banana" level use case, but for general agents (AI image editing gave both OpenAI and Google their viral user acquisition moments it seems). So we're talking something that can go truly viral... Maybe like the AI completely redesigning the entire UI of someone's computer like Mirage via a generative UI... could be very cool!