Choosing between a Claude subscription and Codex for your coding agent workflow right now? The fight has shifted fast, and the winner isn’t who you’d think six months ago.
Matthew Berman just dropped a breakdown that walks through every Anthropic stumble of the last two months, and it’s painful to watch if you’re a power user. The creator pays $200 a month himself and says Anthropic is actively making it harder to spend the tokens he’s already bought.
Here’s the core of what the original poster argues. Dario made one big call 18 months ago. He decided not to bet the company on massive compute capex. Sam Altman made the opposite bet. Demand exploded past every projection, and now Anthropic can’t serve its own flywheel. OpenAI can.
The criteria that matter for picking a side
- Reliability of quota (can you actually use what you pay for?)
- Clarity of terms (is third-party harness use allowed?)
- Uptime (your agents stop when the API stops)
- Model quality for agentic orchestration
How they compare right now
🔴 Anthropic (Claude Pro/Max)
Pros: Opus 4.6 and 4.7 still top-tier for agentic orchestration and tool calling
Cons per the author: Claude Code removed from some Pro tiers in a 2% test, OpenClaude usage restricted, peak-hour session limits tightened, API uptime sitting near 99% (barely one nine past the decimal), Opus 4.7 burns 1 to 1.35x more tokens due to a new tokenizer plus more thinking tokens
🟢 OpenAI (Codex)
Pros: 99.9%+ API uptime, repeated quota resets announced publicly, Codex works inside OpenClaude with no policy whiplash, Peter Steinberger (OpenClaude creator) joined OpenAI
Cons: Codex still trails Opus slightly inside OpenClaude per the creator’s testing
The author’s take
The expert isn’t saying Claude models are bad. He uses Opus every day. His point is trust. Anthropic pushes policy changes on Friday afternoons before holiday weekends, promises clarification, then never delivers it. Two months after the first OpenClaude confusion, the docs still don’t spell out what’s allowed. The creator calls this the real product problem: not the model, the vendor relationship.
Meanwhile Codex keeps resetting limits for any reason. Tibo from the Codex team literally tweeted that transparency and trust are principles they won’t break even if it means earning less. The contrast is brutal.
Practical steps if you’re migrating or hedging
- Keep Opus as your primary orchestrator inside OpenClaude if quality matters more than policy risk
- Add Codex as your async workhorse for long-running background jobs
- Shift token-heavy work to off-peak hours on Claude (weekdays 5am to 11am Pacific burns session limits fastest)
- Watch the AWS Trainium deal timing. Anthropic committed $100B+ over 10 years, but new capacity lands months out
- Treat third-party harness access as unstable. Don’t build a business around it
The deeper lesson from this industry pro
When a company sells tokens at a loss and runs out of compute, they can’t raise prices without losing users to a competitor who can serve demand. So they introduce quiet restrictions instead of honest price hikes. The person who shared this calls these half-measures, and once you see the pattern you can’t unsee it.
Check the full video for the compute-vs-demand spectrum breakdown across Google, OpenAI, Anthropic, and xAI. The Google section alone is worth it.