⚠️ Railway users: rotate API keys; all users: secure your accounts — Following Railway’s environment-variable security incident, revoke and replace unrotated keys stored there and update deployments. Enable APIYI 2FA or a passkey, cap test-key spending, and allowlist stable public egress IPs for production keys.📖 View details
⏸ OpenAI retired 12 legacy model entries and five additional fine-tuned model categories on October 23. Please migrate promptly — Affected entries include specified GPT-3.5, GPT-4, GPT-4.1 Nano, GPT-4o, o-series, and GPT Image snapshots and aliases; check applications and workflows for old references, update
model, and validate behavior.📖 View details📖 New Token Management API scenario: a daily budget per key that resets the next day — There is no built-in daily quota, but a scheduled job at 00:00 (UTC+8) that sets each key’s
remain_quota back to its daily cap does it; keys don’t change, calls stop without charge once the day’s budget is used, and they work again the next day. A ready-to-run script is included.📖 View details🚀 The community ComfyUI node Luck Nano Banana Pro now supports Nano Banana 2.1 — After the author’s update, new nodes default to
gemini-nano-banana-2.1, with NB2 / Pro / 2.5 kept; existing workflows do not switch automatically, just pick it in the model dropdown.📖 View details✅ Gemini is stable again and can be used as usual — The intermittent fluctuations on Google’s official Vertex resources this afternoon have recovered and calls are working normally. We will keep monitoring.📖 View details
⚠️ Gemini is intermittently unstable today, mainly due to Google’s official Vertex resources — It looks like a performance drop on Google’s side, coming in waves rather than a continuous outage; rumors link it to compute allocation around Gemini 4, with no official statement yet. We are monitoring.📖 View details
📖 New FAQ: why the Hermes model list shows only some models — The picker shows at most 50 models per provider; that is a menu limit, not a token permission limit. Switch to other models directly with
/model <full model ID>.📖 View details📖 New FAQ: no Files API; send files inline — This is why
/v1/files returns no available channels for model, unrelated to your group or balance. Assistants and Batch are not provided either; put images and PDFs directly into the request as base64 or a URL.📖 View details📖 New FAQ: 4K images never arrive over a direct connection — With the proxy bypassed, 1K/2K work but 4K hangs with no error, and the log shows
write_response_body_failed with no charge; the direct line cannot carry responses of tens of MB, so route image requests through the proxy.📖 View details📖 New FAQ: OpenAI Fast / Ultrafast mode is not yet supported — Requests with
service_tier (fast / ultrafast / priority / flex) do not error; they run and bill at the standard tier, /v1/responses returns default, and nothing extra is charged.📖 View details📖 For earlier updates, visit the Live Updates Archive — browse history by month, category, or vendor.
Deep Dive
AI Radar
Telegram
Global users
Media Models
New additions