What's new

A service line is retiring a batch of older models on 2026-10-09, so we removed them ahead of time rather than let calls start failing. Affected ids include doubao-seed-1.8 / 1.6-* / doubao1.5-lite-32k, glm-4.7-flashx, qwen3-32b, qwen3-vl-*, qwen-turbo, qwen-vl-plus, and the deepseek-v3.x and R1-distill families. Recommended replacements: doubao-seed-2.0-pro or 2.1-pro, glm-5.3-flash, qwen3.7-plus, deepseek-v4.1-flash.

deepseek-v3.1, glm-4.7, qwen3-max, qwen3-coder-plus and qwen3-coder-next keep working — they are served over another line now. Their cached-input price rises (deepseek-v3.1: $0.12 → $0.603 per 1M), and glm-4.7 input goes from $0.45 to $0.63 per 1M. Input and output prices for the other three move by under 1%.

Heavy cache users on deepseek-v3.1 should consider deepseek-v4.1-flash: cheaper per token ($0.30 / $1.20) and cached input at $0.006 per 1M.

New: the full Doubao Seed lineup, GLM-5.3 / GLM-5.3-flash, MiniMax-M3, DeepSeek V4.1 Flash, and Qwen3.8 Flash.

The GLM line got cheaper by roughly 30–43% (e.g. glm-5.1 input: $1.47 → $0.90/M).

Some models now price by input length in tiers.

Streaming responses could end prematurely when the first token — or a mid-stream pause — took more than 2 seconds. This is now fixed.

When a service line fails before the first chunk is sent, streaming requests now automatically fail over to a backup line.

Balance top-ups on iOS now go through Apple's In-App Purchase, per App Store Guideline 3.1.1 — required for any app that lets users buy digital credit for use inside the app.

Docs are one click away from anywhere in the console instead of buried in the sidebar.

The EN/中 switch in the header previously reset on some bare routes (privacy, pricing). It now persists across the whole site, and the privacy policy is fully bilingual.

Every model is now priced at our wholesale cost + 5% (just covers payment fees) instead of a markup — DeepSeek is the one exception, kept at a flat 8折 (80% of list) as the headline value tier.

Follow-up fix the next day: the English pricing pages were displaying rates per 1K tokens while every other page used the industry-standard per-1M convention — same underlying price, just a confusing unit. Now consistent everywhere.

Android users can now download the app directly from the marketing homepage, no Play Store detour.

Bare-metal/VM instances gained disk & instance snapshots, capture-as-template custom images, attachable data disks, per-tenant VPC isolation, and an EIP bandwidth-adjust dialog — all proxy-billed with proration.