#OpenAIRemovesCodex5HourLimit

The AI Infrastructure War Just Escalated and It's Not About the Models Anymore

Last week the battleground shifted. Not in the labs where researchers fine-tune the next frontier models. Not in the benchmark leaderboards where Sol edges out Fable by a few percentage points. The real fight broke out in the billing dashboards.

OpenAI blinked first. Then within the hour Anthropic blinked back.

Here's what actually happened: OpenAI quietly removed the 5-hour usage cap on Codex and ChatGPT Work for Plus Pro and Business subscribers. Just... gone. A hard limit that had developers rationing their prompts like water in a drought suddenly evaporated. The company also announced an immediate quota reset for its 6 million active Codex users yes six million people now treat AI coding agents as essential infrastructure.

Anthropic's response came fast. Claude Fable 5 access which was supposed to expire July 12 got extended to July 19. Then extended again. The pattern is telling: every time OpenAI makes a move on access, Anthropic finds another week.

But this isn't generosity. This is competitive panic dressed up as customer appreciation.

Behind these moves sits an uncomfortable truth the AI labs don't advertise: they're all compute-constrained. Every single one of them. OpenAI, Anthropic, Google, Meta they're racing to secure gigawatts of data center capacity while pretending they have infinite supply.

OpenAI's internal memo to shareholders (leaked to CNBC in April) revealed the scale of the anxiety. The company projected Anthropic would have 7-8 gigawatts of compute by end of 2027. OpenAI? They're targeting 30 gigawatts by 2030. That's not confidence. That's an arms race where the ammunition is Nvidia chips and electricity contracts.

When you're compute-constrained, you have two levers: raise prices and alienate users, or eat the margin and pray your infrastructure catches up. OpenAI just chose door number two for now.

The 5-hour Codex limit was becoming a retention disaster. Developer forums were filling with cancellation stories. "I hit the limit halfway through a task," one user wrote. "Ten times less productive than last week." Another: "Already cancelling Plus and moving to a third-party provider. Same price, double the usage, no limits."

In the AI race, limits are churn accelerants. Developers don't forgive friction when alternatives exist. Anthropic learned this with Fable 5's rollout their strict guardrails triggered a backlash so severe that Business Insider reported the company admitted it "made the wrong tradeoff." Users were sharing screenshots of massive credit bills and cancellation confirmations.

The Efficiency Gambit

OpenAI isn't just removing limits; they're optimizing consumption. GPT-5.6 Sol now consumes 54% fewer tokens on agentic coding tasks, according to Sam Altman's comments to CNBC. That's not a vanity metric. That's a direct response to enterprise anxiety about AI spend.

Every CTO in America is asking the same question: "What are we actually getting for our AI budget?" OpenAI's answer is efficiency. Anthropic's answer is... well, they're still figuring that out. Fable 5's pricing post-promotion $10 per million input tokens, $50 per million output tokens makes it the most expensive generally available model on the market.

Here's the uncomfortable prediction: these extensions and temporary removals are just that. Temporary. Anthropic has already signaled Fable 5 will transition to a credit-based model after the promotional window. OpenAI hasn't committed to keeping the 5-hour limit gone forever.

The pattern is clear. Both companies are using access as a competitive weapon while they build out infrastructure. The goal isn't to give away compute indefinitely. It's to make you dependent before they start charging properly for it.

The Real Battleground

Forget model benchmarks for a moment. The next phase of the AI war isn't about which model scores higher on coding evaluations. It's about who can deliver reliable, unlimited access at a price that doesn't trigger CFO panic attacks.

Meta just entered the fray with Muse Spark 1.1, pricing it "very aggressive and attractive" compared to OpenAI and Anthropic. They're limiting API access to their own properties for now, but the message is received: there's room to undercut the incumbents.

The infrastructure economics are brutal. Anthropic is reportedly paying xAI $1.25 billion per month through 2029 for Colossus compute access. That's $15 billion a year for one data center relationship. These are telecom-level infrastructure commitments, not software margins.

If you're building on AI right now, you're living in a golden window. Unlimited Codex. Extended Fable access. Aggressive pricing from challengers. But windows close.

The smart move isn't to assume this lasts. It's to build systems that can switch models without rewriting your stack. The labs want lock-in; your defense is abstraction. Treat these models like commodities because, eventually, they will be.

The 6 million Codex users aren't just a user metric. They're a warning shot. When that many developers embed a tool into their daily workflow, removing limits isn't charity it's market capture. OpenAI and Anthropic aren't fighting for today's subscription revenue. They're fighting to own the infrastructure layer of the next decade of software development.

The limits will return. The pricing will normalize. The question is: whose workflow will be too entrenched to leave by then?