🚀 Kimi K3 is live — $3/$15 per 1M tokens, identical to official pricing —— Moonshot AI’s 2.8-trillion-parameter open-source flagship released 7/16, the largest open-source model ever: a 1M context priced flat across the whole range, #1 on LMArena Frontend Code Arena, native image/video input; APIYI has launched
kimi-k3, with cache-hit input as low as $0.30 and recharge bonuses giving an easy ~9% off (recharge $100, get 10%), up to ~17% off — deep dive at Kimi K3 Launch.📖 View details✅ Nano Banana Pro & 2 run on pure official AI Studio relay — consistently stable, bring your traffic —— Since Nano Banana launched in November 2025,
gemini-3-pro-image / gemini-3.1-flash-image and the lite series have run exclusively on the official AI Studio relay (an official Vertex group is also available) — never a reverse-engineered channel; a price well below official does not mean a reverse-engineered source. Professionalism earns trust.📖 View details✅
gpt-image-2-all speed is back to normal — fast cases complete in about 40 seconds —— The ChatGPT website has recovered its compute capacity and this reverse-engineered model’s speed has returned to normal with it, with first-byte times mostly in the 30-60 second range in our logs; send traffic with confidence — no RPM cap to worry about, and supply is ample.📖 View details📊 Nano Banana Pro & 2 have stayed fast for the past 20 days — averaging about 30 seconds per image —— Cards on the table:
gemini-3-pro-image and gemini-3.1-flash-image, the core workhorse image models on APIYI, have been consistently fast in our monitoring over the past 20 days and just as stable under high concurrency; we’ll keep working hard to safeguard operations and supply.📖 View details📚 New docs group “Image API Essentials” with the new “Must-Read & Best Practices” page now live —— Key takeaway: all image models are synchronous calls with no async task ID — a dropped connection loses the result but is still billed, so a generous
timeout is rule number one; the page includes per-model timeout recommendations and a base64 / URL output reference table, and the existing “Image Compression & Output Resolution” and “How to Generate Satisfying Images” pages have moved into this group — read it at Must-Read & Best Practices.📖 View details✅
gpt-image-2-all speed is back to normal — the 7/13 temporary redirect to gpt-image-2-vip has been reverted —— The two models run on separate underlying routes (ChatGPT vs Codex) and remain split by model; gpt-image-2-all generates faster, while gpt-image-2-vip currently supports the size parameter — pick as needed.📖 View details📚 Codex setup method upgraded: one
experimental_bearer_token line works across all three surfaces —— The old env_key form only reads the launching process’s environment variables, so the desktop app can’t see terminal exports and often fails with Missing environment variable: OPENAI_API_KEY; the new form puts the key directly in config.toml, and one config now covers the desktop app, IDE extension, and CLI — cause and fix in the Codex integration guide.📖 View details⚠️
gpt-image-2-all is timing out — the platform has temporarily pointed it to gpt-image-2-vip —— Caused by an image-generation anomaly currently showing on the ChatGPT website; this reverse-engineered model’s availability tracks the official site. This is a proactive safeguard to protect production traffic: the interface stays unchanged and fully compatible, pricing is unaffected, and it will be switched back once the website recovers.📖 View details📚 New docs section “Grok API Calls” — 5 hands-on-verified bilingual pages now live —— Covering pricing and the capability matrix for the full
grok-4.5 / grok-4.3 / grok-4.20 / grok-build-0.1 lineup, chat and reasoning, web and X search, code execution and Remote MCP, plus the multi-agent model; every conclusion verified against 56 test requests — start at the Grok Overview.📖 View details✅
gpt-image-2-all speed is back to normal, and the timeout credit reissue is complete: $850 total to 540 users —— Last week’s ChatGPT website compute issues intermittently slowed this reverse-engineered model; we exported 14,500 billed image logs that exceeded 300 seconds and reissued credits one by one, rounding amounts under $1 up to $1. Check the top-nav “Recharge” page for the record (type: Admin). The reissue isn’t the goal — stability is, and we’ll keep working on operations and supply.📖 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