Большое обновление AI-моделей: свои провайдеры, новый BYOK, ссылки для приложений

Several big changes shipped this week. We've put them all into one update post.

Custom OpenAI-Compatible Providers. You can now connect any OpenAI-compatible service to the platform. Pre-configured: Minimax, Yi, Qwen, Cerebras, Cohere. For everything else — a universal "Custom OpenAI-Compatible" provider with a Base URL field. The platform fetches the model list automatically, or add modelId manually.

AI Models catalog. The section has been redesigned — platform models (BitrixGPT 5, GPT-OSS 120B, Gemma 4 26B, Gemma 4 26B Thinking, Whisper, Vision) are now front and center. Free through the platform, all you need is a Vibe API key. Each model card has a "Use" button — click to copy a ready-made prompt for an AI agent. No need to remember how to call the model — just paste and go. Take a look: https://vibecode.bitrix24.tech/ai

BYOK moved to a dedicated section. Your external provider API keys now live in a separate section with cards. Each key shows its status (valid or not), period stats, and a Verify button. Deletion via a proper confirmation dialog. You can add multiple keys — the platform will tell you how to use a model with each one.

Vision via standard OpenAI format. Image support in models works via the standard OpenAI Vision format: content is passed as an array with objects of type: text and type: image_url. Image source — data:image/...;base64,... or https://.... Models with Vision: bitrix/bitrixgpt-5-vl, anthropic/claude-opus-4-6, anthropic/claude-sonnet-4-6, anthropic/claude-haiku-4-5-20251001. For Anthropic, the platform converts the request to the native format automatically — nothing extra to do.

Public links for Black Hole apps. Black Hole servers now have public links for applications. Previously, giving someone access required setting up permissions. Now you can just send a link to a person or chat — no extra setup. Handy for demos, presales, tests with colleagues and clients.

What's next. In progress — Markdown output for models, trial system improvements, AI Models section revision toward a catalog concept. If something doesn't work or you want a feature — send feedback right from the app through the AI agent, we'll look into it.