OpenAI boosts ChatGPT Plus limits and adds UI model indicator

Nothing frustrates paying AI power users faster than a tool that suddenly feels dumber, especially when they cannot tell which model is actually answering behind the scenes.

OpenAI CEO Sam Altman just stepped in to calm nerves following the launch of GPT-5, announcing that the company is significantly increasing rate limits for reasoning capabilities for ChatGPT Plus subscribers. According to Altman, usage caps across all model classes will soon exceed where they stood before GPT-5 rolled out.

What went wrong under the hood

The quick adjustments follow an admittedly bumpy rollout. OpenAI acknowledged that dropping multiple massive updates at the same time caused unexpected turbulence, leading to widespread user confusion.

The biggest culprit was a temporary failure in the autoswitcher. GPT-5 was designed as a unified system that routes prompts between standard, mini, and nano model sizes based on complexity, removing the need for users to choose manually. When that routing mechanism broke, the model appeared significantly less capable than it actually is.

To restore transparency, OpenAI is preparing a UI update that clearly reveals which specific model is handling each response. The team is also planning a manual trigger so users can force advanced reasoning on demand, while exploring options to let Plus users keep accessing GPT-4o.

Reasoning is becoming the default

Surging user demand left OpenAI with little choice but to bump the limits. Daily usage of reasoning models has exploded across the board.

Among free tier users, daily reasoning engagement jumped from under 1% to 7%. Among Plus subscribers, it leaped from 7% to 24%. With people leaning harder into complex problem-solving, restrictive caps were quickly becoming a bottleneck.

When operating properly, GPT-5 packs serious muscle. It hits 74.9% on SWE-bench Verified and 88% on Aider polyglot, outperforming GPT-4o while slashing hallucinations by 45%. It also brings integrations with Gmail and Google Calendar, alongside four distinct persona options: cynic, robot, listener, and nerd. However, rolling out changes at this scale is taking longer than initially anticipated.

Why it matters

For Plus subscribers, control and predictability matter just as much as raw benchmark scores. Automated model routing makes sense from an infrastructure and cost perspective, but black-box switching breeds immediate skepticism whenever output quality dips. Letting users see which engine is running and manually invoke deep reasoning restores essential trust.

For the broader AI industry, this episode highlights the risks of building complex routing layers on top of frontier models. As AI architectures shift toward compound systems that dynamically balance cheap and heavy models, orchestration becomes the ultimate point of failure. Delivering peak intelligence is no longer enough; keeping the pipeline transparent to users is just as critical.