OpenAI leaks hint at GPT-4 cap removal and DALL-E 3 inpainting

If you use ChatGPT on a daily basis, running into the 40 messages every 3 hours cap is a familiar frustration. The same goes for DALL-E 3, where changing a tiny mistake in an otherwise great picture usually means starting all over again.

OpenAI may finally be addressing these pain points. Developer Tibor Blaho, who has previously uncovered upcoming ChatGPT features in code, found several major additions currently in testing for GPT-4 and DALL-E 3.

Smarter routing and flexible caps

The most noticeable tweak is the potential end of the rigid message limit. In a preview version of ChatGPT, Blaho spotted a model toggle that mentions a possible limit rather than the strict 40 messages per 3 hours rule.

Alongside this, OpenAI is testing a Model Tuner Selector. This system automatically splits prompts between GPT-3.5 and GPT-4 based on complexity. Simple questions get handled fast by GPT-3.5, while complex problems route to GPT-4, saving users their high-tier quota. The routing can be turned off in chat settings.

For free-tier users, an upgraded response button is also in testing, allowing them to re-run a weak GPT-3.5 answer through GPT-4 with a single tap. OpenAI is also preparing an anonymous usage option, letting people test ChatGPT without creating an account.

DALL-E 3 gets an editor

Image generation is finally getting the update users have asked for since launch. Blaho shared previews of an inpainting editor inside ChatGPT, allowing users to highlight a specific area of an image and describe edits in text.

The interface is shifting to support this workflow. The chat moves to a side panel, keeping the image and the editor front and center.

Blaho has a strong track record, having revealed both the memory feature and multi-GPT mentions before OpenAI officially announced them.

Why it matters

For paying subscribers, an inpainting canvas finally makes DALL-E 3 useful for real work. Up until now, Midjourney and Ideogram have outpaced OpenAI in fine-grained control, especially embarrassing given that DALL-E 2 had inpainting from early on.

For OpenAI's business, these features serve a double purpose. Removing friction with anonymous chats attracts fresh users, while the one-click upgrade button gives free users a direct taste of why GPT-4 is worth 20 dollars a month.

We think the smart routing feature is clever for managing backend compute load, but it comes with real risk. If the selector misjudges a subtle prompt and hands it off to GPT-3.5, paying users will quickly accuse OpenAI of stealthily degrading model intelligence. Keeping an off switch in settings is a necessary safeguard.