⚠️ The
gpt-image-2.5-vip series is at concurrency capacity; we are adding resources — gpt-image-2-vip is mostly normal and gpt-image-2.5-all is stable; meanwhile you can use the 2nd generation or the official gpt-image-2.5-sunburst. Reverse channels depend on provider-side supply and risk controls, so keep multi-model fallbacks on the client side.📖 View details📚 New FAQ: why logs show Haiku when Claude Code is set to Opus — it’s the client’s own cost-saving behavior: background tasks like session titles and subagents like Explore default to
claude-haiku-4-5-20251001 and are billed at Haiku pricing. To run everything on Opus, point the related environment variables at claude-opus-5-5.📖 View details✅
gpt-image-2-all / gpt-image-2.5-all are back to fast, stable generation now that OpenAI has recovered —— gpt-image-2-vip and the 2.5 -vip series occasionally hit concurrency shortfalls, with good overall success rates; the official gpt-image-2 / gpt-image-2.5 series are working normally.📖 View details⏸
gpt-image-2 no longer supports transparent backgrounds: OpenAI shut down its background: "transparent" on 2026-09-30, and it now returns 400 (not billed) —— for transparent images use the official gpt-image-2.5-flare (text-to-image) or gpt-image-2.5-sunburst (edits); all three share price and parameters, so only the model name changes. The transparent background FAQ has been updated.📖 View details📚 New FAQ separates the two Gemini thought-signature 400s in tool calls:
missing a thought_signature means the client dropped the signature when replaying history and must fix its replay logic; Invalid / Corrupted thought signature means the current route does not accept it, so resend as is, then strip history signatures or start a new conversation —— Gemini 3 and later require the functionCall signature back unchanged; on /v1/chat/completions the gateway handles it for you.📖 View details⚠️
gpt-image-2-vip is under heavy load — when reverse-engineered routes fail, fall back to the official models first —— -vip requests surged today and demand is outpacing supply (a demand spike, not a capacity cut); gpt-image-2.5-sunburst-vip runs on separate capacity and is doing better; gpt-image-2-all is degraded by the OpenAI provider-side outage; the official gpt-image-2 / gpt-image-2.5 series are working normally.📖 View details⚠️ Official DeepSeek models are timing out today due to a provider-side DeepSeek outage; APIYI has urgently enabled a Volcengine full-size route, performing normally —— DeepSeek’s status page shows a partial Web/API disruption lasting about 3 hours;
deepseek-v4-pro mostly times out while the Flash series is less affected. Model names and calling conventions are unchanged; please report any remaining issues.📖 View details⚠️
gpt-image-2-all / gpt-image-2.5-all are returning model not found and 429 insufficient-concurrency errors today, caused by a ChatGPT web outage —— we are working on it; the official gpt-image-2 / gpt-image-2.5 series is working normally and the -vip series (e.g. gpt-image-2.5-sunburst-vip) is currently holding up. Try them online at imagen.apiyi.com.📖 View details📚
apiyi says command not found after npx apiyi@latest check? Keep the npx apiyi@latest prefix on later commands —— npx runs the CLI without installing it; install once with npm i -g apiyi to use plain apiyi. A new FAQ covers the three ways to run it, and from CLI 0.2.2 the hints print the full command.📖 View details🔧
oxygen-1.0 parameter update: resolutions are now 480p / 768p, and 320p is retired — 768p output is now 1376×768; the envelope gains optional quality (quality tier) and prompt_enhance (prompt enhancement) parameters. Default behavior is unchanged and the price stays at $0.02 per second.📖 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