Yu Pan

The quality jump is real — outputs feel more intent-aware and less like prompt guessing
• Speed + reliability makes it usable for daily, production-level workflows, not just demos
• The ecosystem effect is huge: devs, creators, and teams can all build on the same foundation
• It’s one of the few AI products that keeps improving without increasing cognitive load

What’s most impressive is how OpenAI continues to turn cutting-edge research into something immediately practical. Curious what you’re most excited to unlock next with this release

• More transparency and control around model behavior and updates, especially for teams using it in production
• Clearer guidance on best practices across different use cases (dev, design, marketing, ops)
• Better long-term memory / project-level context to reduce re-explaining complex systems
• More predictable pricing and usage limits as capabilities continue to expand

Still an incredible product — these improvements would make it even easier to rely on at scale.

We chose OpenAI because it consistently strikes the best balance between capability, reliability, and developer experience. The models are strong across reasoning, multimodality, and real-world tasks, but what really stands out is how quickly those advances become usable products.

Beyond model quality, the ecosystem matters: stable APIs, clear documentation, and a fast-moving community make it easier to go from prototype to production. Compared to alternatives, OpenAI feels less like a single model and more like a long-term platform we can confidently build on.

Overall reliability has been solid, even during major launches. That said, I’m curious how the team is thinking about model uptime and latency during peak traffic — especially as more teams depend on it for production and customer-facing workflows.

Realtime Voice is promising for conversational and agent use cases. Curious how it performs over SIP/WebRTC in real-world conditions — especially around latency, jitter handling, and stability during longer sessions.

Image input is already quite strong, especially for general visual understanding. Curious how robust it is for more demanding OCR and visual QA tasks (e.g. dense tables, low-quality scans, or mixed-language documents), and what improvements are planned there.