Updated November 12, 2026: Qwen Code's legacy free OAuth allowance ended on April 15, 2026. Our previous recommendation to start with 1,000 free daily OAuth calls was out of date. The current Qwen Account with Subscription route requires an active Plus or Pro plan. This correction preserves the original article URL and publication date.
The budgeting question is not simply whether the CLI is free to install. It is which service pays for the model behind each coding task. Separate the client, authentication method, inference provider and plan before comparing prices.
Is Qwen Code still free?
The official terms notice distinguishes today's subscription account access from the discontinued free allowance. Open-source software does not imply free hosted inference. A third-party provider's promotion or a locally hosted model is also not a restoration of the old Qwen OAuth quota.
The current access choices
Qwen's authentication documentation lists subscription sign-in, QwenCloud, Alibaba ModelStudio, third-party providers and custom providers. Plus/Pro account access uses OAuth. Provider-based routes may use a plan or metered API billing; their endpoints and entitlements are not interchangeable. For noninteractive environments, the documentation directs users to API-key authentication.
Do not copy a price from one regional provider into another provider's budget. Check the selected model, account region, billing currency, quota period and any supported-tool restrictions in the account that will actually run the work. We are not quoting one universal monthly Qwen Code price because these are different purchases.
Choose a route by workload, not by the word “free”
| Situation | What to evaluate | Main budgeting risk |
|---|---|---|
| Individual interactive coding | Current subscription entitlement versus a small API trial | Buying a plan whose allowance does not fit the workload |
| Unattended repository tasks | An approved API route with scoped credentials | Retries and parallel jobs multiplying spend |
| Team rollout | Provider access, spending controls and account ownership | Confusing pooled credits with a guaranteed per-person allocation |
| Local model experiment | Hardware, model compatibility and task quality | Treating local compute and maintenance as costless |
Measure cost per accepted change
Nerova's recommendation is to evaluate a small set of representative tasks before choosing a larger plan. Include a straightforward edit, a repository investigation, a bug with tests and a task requiring several tool calls. Use the same acceptance criteria across options and retain failures rather than dropping them from the comparison.
Record total provider charges, completed tasks, reviewer minutes and retries. Divide the full trial cost by the number of changes a reviewer accepts. A cheap first response is not necessarily a cheap completed task.
For example, a hypothetical $12 trial producing six accepted changes costs $2 per accepted change before human review. A $9 trial producing only three costs $3. These are illustrative calculations, not measured Qwen performance or quoted provider prices. Include review time separately if one option creates substantially more cleanup.
Guardrails before expanding usage
- Set a trial spending limit and inspect actual provider usage after each batch.
- Keep credentials out of repositories and use a separate trial project where available.
- Limit parallel tasks and stop runaway retries instead of assuming a subscription makes every action free.
- Check the provider's data terms before submitting private source code.
- Repeat the budget estimate when changing the model, context size or authentication route.
Keep the evaluation window long enough to include normal work, but define the stopping rule in advance. Stop early for unacceptable data handling or uncontrolled spending; do not declare a winner from one easy prompt.
Bottom line
Do not plan around the retired free OAuth allowance. Compare today's paid account access with the specific provider or local route you can operate, then decide using accepted output, review effort and total cost. The lowest advertised entry price is only useful when the resulting workflow reliably does the job.