Chai AI

Most people who want to chat with an AI character end up on Character AI, and most people who want a general assistant open ChatGPT. Chai AI says it got there before both. According to its own timeline, CHAI launched on the App Store in November 2022 as a consumer platform for creating your own ChatAIs, "ahead of Character AI and ChatGPT." The difference in purpose is the whole story. ChatGPT is built to answer questions and finish tasks. Chai is built for one thing: letting you create and talk to your own AI characters and stories. It isn't trying to write your emails or debug your code.

The gap shows up most clearly in how Chai trains its models. The company's roadmap reads like an engagement playbook. It moved to an in-house 6B model in February 2023 because open-source LLMs "no longer satisfied" its social use case, then scaled to 13B, 24B and, by December 2025, 235B A22B Mixture-of-Experts models. Along the way it trained reward models directly on user signals (100 million of them for Reward Model XL), used PPO "to decrease the probability a chat session ends," and added DPO and GRPO for reported engagement gains of 20% and 15%. It also claims to have invented "model blending," which ensembles different LLMs trained on different targets at the conversation level. The infrastructure behind it is real: 5,000 GPUs across Nvidia A100, L40S and AMD MI325X and MI300X chips, 1.2T tokens served per day and 51K unique LLMs in production.

Here's the honest read. Chai is unusually open about what it optimizes for: retention, screentime and revenue. It credits its December 2025 upgrade with a 25% lift in both screentime and revenue. That's good news if you want characters that keep a conversation going, and worth noting if you're wary of apps designed to hold your attention. It's listed as freemium, with paid plans starting at $13.99/mo. The public site at chai-research.com is mostly a research, funding and careers story ($55M+ raised, backing from AMD Ventures and CoreWeave, 10M+ active users), so you'll learn far more about its GPU cluster than about the app's actual features before you install it.