Qwen Code Live and batch jobs: API keys, Chrome support and cost limits

Choose the entry point you actually need

Qwen Team's October 9 weekly update describes these capabilities as available: voice collaboration through Live, working in Chrome with your existing login, sending tasks by email, and background /batch-api jobs. This guide follows the current usage documentation. The update spans several releases, so its coverage does not establish what your installed version supports.

For help with an error in front of you, start with Live. For a browser page, check the Chrome extension next. Email Channel is useful when you want to assign work to your computer while away. Choose batch processing for many independent files; keep tasks that need several rounds of discussion in a normal conversation.

Enable Live with a dedicated Realtime API key

In Web Shell or the desktop app, click Live in the left sidebar, open the Qwen Live card on the right, turn it on, enter your Bailian DashScope Realtime API key and click Save. Realtime is the live voice service. An ordinary conversation login, Qwen OAuth or another model's key does not establish access to it.

After saving, select Open Live Voice on the Live page or use the realtime voice button at the bottom right of the composer. Enable screen sharing during the call only when you need the assistant to see the screen. Move passwords, private messages and unrelated windows out of view first.

For a hypothetical example, ask a voice question about a test project's error before deciding whether to share its window. An available button does not prove that the account has quota. Check Realtime pricing, regional availability and account eligibility for the service you use; the weekly update does not establish free access.

Chrome keeps your login; this feature still excludes Windows

Follow the weekly update's link to Qwen Code in the Chrome Web Store and verify the Qwen Team publisher. Current Browser Use documentation requires Chrome 125+ and Qwen Code 0.24.2+; the store listing also requires Node.js 22 or newer. Existing-session Chrome control currently supports Mac and Linux, with Windows not yet supported.

Keep only one Qwen Code extension in the target Chrome profile. Disable or remove an older unpacked extension before following the Browser Use instructions to connect the local client. Versions, extension installation and the connection procedure are separate requirements; a successful store installation does not prove that the local connection works.

Once connected, start with a clearly scoped read-only task, such as summarizing an open test report. The assistant can use existing tabs and sign-in sessions, but it can also fill and submit forms. Define the destination and permitted actions before assigning a task that writes or submits anything. A signed-in session does not authorize every possible action.

The extension listing says page content, URLs, screenshots, browser-history results and other task outputs can reach local Qwen Code and be included in requests to the configured model provider. Background operation does not mean local-only data processing. Optional Web Shell browser automation also requires a separately configured Chrome DevTools MCP adapter; do not treat it as the same already-configured entry point.

Use a dedicated mailbox and understand the allowlist

Enable IMAP and SMTP on a dedicated mailbox: IMAP receives tasks, while SMTP sends replies. Follow the official Email documentation to add an email-type channel in settings.json with the address, mail servers and working directory. Reference password environment variables from the running process rather than putting plaintext passwords in the configuration.

Keep privatePolicy set to allowlist, add your bare email address to allowedUsers, and run qwen channel start agent-mail. agent-mail is the channel name in the official example; use the name in your actual configuration. The first connection skips existing mail in the selected folder, then later connections resume the saved UID cursor. Deleting the state file is not a retry procedure.

After the channel connects, send a narrowly scoped task from an allowed address and reply to the resulting email to continue. A From allowlist filters addresses; it does not authenticate senders, so the mailbox provider must also filter forged mail. The current documentation excludes provider OAuth, mailbox management, calendar handling and encrypted-mail processing.

If execution stops or SMTP acceptance is uncertain, inspect the mailbox and task side effects before following the recovery instructions. The client preserves unresolved records and refuses automatic replay. Sending the same task again is no substitute for finding out what happened to the first one.

Before /batch-api submits, check credentials, samples and approval

The Batch documentation lets you select batch.model separately in settings.json and use its model entry's baseUrl and envKey for the DashScope API endpoint and key. Your ordinary conversation can keep Qwen OAuth or another provider; OAuth itself has no Batch route. This executor uses OpenAI-compatible chat-completions and does not support wireApi set to responses.

Restart the interactive session after changing the selection, then run qwen batch check to verify readiness without submitting a paid request. Use /batch-api to describe independent work and inspect a small sample, output directory, token estimate and plan. This document-translation example has not been executed: /batch-api translate the Markdown docs in docs/zh into English, writing them to docs/en with the same file names.

Read the dry-run preview before approving submission. It freezes a snapshot of the plan, source files and settings; changes between preview and submission cause refusal. Results wait and collect in the background. Schedule work around the provider's actual completion window rather than treating the weekly update's typical timing as a delivery guarantee.

Half-price requests do not halve every cost or make planning free

The current Batch documentation limits the discount to successful requests at 50% of the realtime list price and says batch requests do not hit the prefix cache. For many short outputs sharing a long prompt, existing realtime cache discounts may be more attractive. Do not move every task to batch solely because of the half-price label.

The weekly update calls the initial plan free. The operating documentation says dry-run uploads nothing and incurs no batch-request billing, but its estimates exclude the ordinary conversation spent preparing the plan. The practical distinction is that an unsubmitted batch has no batch-request charge, while ordinary model turns can still consume tokens. Thinking tokens, actual model prices and account region also affect the total.

After waiting, distinguish delivered, held and failed: written to the target, held because source or target content changed, and a failed request or result. Failed items are not retried automatically because retries may bill again. If a lost submission response leaves submit-unknown, reconcile through collect first instead of submitting again and risking a duplicate charge.