Customer-facing improvements, stability updates, and platform fixes.
Feedback submitted under the Oct 22 terms earns $5 USD or ¥10 CNY when adopted and rewarded. Feedback submitted under earlier terms keeps its original $10 USD or ¥20 CNY reward, even if it is reviewed or granted later. The monthly cap remains five rewards per account.
The first-token latency we show and alert on now stops at the first streamed output of any kind, including a reasoning block, not the first visible answer text. Models that think before answering will read lower; the model and status pages say so.
The site notice banner no longer refreshes while your tab is hidden, and refetches once you come back. Long notices wrap instead of overflowing the strip.
The console Chat page now shows the inbox bell. Video job cards show a recovery line while a job is retried. The homepage Playground's Agent mode has layout and wording fixes, and the sign-in and sign-up pages no longer break Chinese titles mid-word or hide buttons behind phone cookie controls.
The homepage Playground now has two tabs, Chat and Agent. Agent mode opens with example tasks and the tools each one calls. The Tools button lists every tool with a one-line description, and you can switch tools on or off before you run. Trying it is still free.
WeChat accounts can now add an email in Settings and verify it with a code. If you signed in more than 10 minutes ago, scan the WeChat code once to confirm it is you. New WeChat sign-ups can add an email at the end, or skip it.
In an app's built-in browser, such as Facebook's, Google, X and GitHub sign-in now continues in your phone's browser. Android jumps there automatically, and Android and iOS both show an Open in browser button. WeChat blocks this, so a guide appears. You can also copy the link or use email.
The console Playground is now Chat, at /console/chat. Model, system prompt, temperature and other settings sit in a left panel beside the conversation, and cURL export stays. Old Playground links open Chat, and you can try it without signing in. Video has its own page.
The language you choose, 中文 or English, now applies in every tab and on every later visit. Region sets the default language only for visitors who have never picked one. Currency and prices still follow your region.
The launch offer now has no end date: models with a launch price keep billing at it. If prices change, we will give at least 3 days' notice before the change takes effect. Details are on the launch page.
After you sign in with WeChat, your WeChat nickname and avatar appear on your account. If your account has no email yet, the console now suggests adding one for receipts and notifications. You can turn the reminder off.
After you submit an image, the order stays in your history even if you switch pages, reload, or open the console on another device. Its prompt and picture stay with it, and it keeps updating until it finishes.
New accounts created with WeChat open in the China region with a CNY wallet and CNY prices, wherever you sign up from. New users accept the terms after scanning. On a phone, save the sign-in QR code to your photos and scan it in WeChat. Existing accounts are unchanged.
The sign-up and sign-in pages now offer WeChat. Scan the QR code to create an account or sign in, without an email address. Scanning again with the same WeChat account opens the same account. Email, Google, GitHub, and X sign-in are unchanged.
When your balance or project budget cannot cover an image, each account can still generate up to 7 free images per UTC day, in the five standard sizes. It resets at 00:00 UTC. When the 7 are used, the API says so.
inferway/image-gen-v1 (Z-Image Turbo, standard tier) is now live for text-to-image. Call it through POST /v1/images/generations or the Interactions API. The image guide has requests you can run as they are. Per-image pricing is on the Models and Pricing pages.
Model pages now show median first-token latency and output speed from the last 24 hours of real traffic. Models without data yet still show Being measured.
Sign-up with X or another third-party account now works without a verified email. After sign-up, the console reminds you to add and verify one. Email notifications need a verified address.
The console Image page now follows the Video page style. Running jobs show elapsed time and can be cancelled. Each result opens a larger preview with download, copy, and reuse actions. Download links refresh when they expire.
On desktop, the homepage model cards sit two per row with the price near the top. Each card states its free-quota status, and the monthly cost estimate can be expanded.
The Requests page now shows image and video jobs with text requests, newest first. The Video page hides jobs it can identify as images. Media rows show each job's own amount.
Accounts without a verified email now see a reminder in the console to add one. It can be dismissed for the current page. Email notifications need a verified address.
When paid and free orders wait together, the queue serves five paid orders for every free order. If no paid order is waiting, free orders start right away. Paid orders wait less, free orders keep generating, and request parameters are unchanged.
A bell in the console top bar shows unread counts. The panel lets you read notices, mark one or all as read, load more items, and closes with Escape or an outside click, including on phones. Account and payment notices appear here.
Video orders using 1 to 3 reference images and up to 158 frames now use a faster generation path by default. Text-to-video, first-and-last-frame modes, and longer clips remain unchanged. Existing API parameters do not need to change.
Video cards and the detail drawer now show a preview image. The drawer also uses it as the player poster before playback.
The console video page puts prompt, mode, duration, clip count, and aspect ratio into one box at the top. Each run creates 1 to 4 landscape or portrait clips with fixed seed support. Cards open a detail drawer, and history can be played directly and grouped by batch.
The gateway now tells a busy engine from an unavailable one, so a slow health reading no longer turns into a 503. Stuck streams are now bounded. The aim is fewer false errors under heavy load.
When output gets stuck in a repetition loop, including inside reasoning output, the gateway detects it earlier. The stream ends early with a clear model_output_repetition error instead of running until hitting the maximum length limit.
The console is lighter on first load, and rarely used panels load only when you open them.
Usage charts now include value axes. The 7-day view labels every day while the 30-day view labels every 7 days, and input and output use two colors. The time-to-first-token chart marks days without measurements as no data.
If the jump from sign-up to the console stalls on a phone, the page refreshes the session and retries before offering sign-out. The form preserves tap target sizes, the cookie notice becomes a slim bar on phones, and the invite-code entry is more compact.
The top-up dialog uses larger type, and preset cards show estimated credit up front with bonus amounts marked by outlined tags. Summary details, inputs, and buttons are larger, and presets display in two columns on phones. The payment flow remains unchanged.