Qwen API Pricing 2026

Alibaba region, promotions and long-context billing

Alibaba rates here are USD per 1M tokens in the Singapore / International scope, not China-region prices. Calls are pay-as-you-go by default. The total input-token count selects the request tier, and all tokens in that request use the corresponding tier rate. Qwen3.7-Plus applies the explicitly published 20% alias promotion; its dated snapshot does not automatically inherit that promotion.

• Qwen3.6-Flash: Standard rate through 256K input tokens; higher request tier above that. Above 256,000 input tokens, the rate is $1 input and $4 output per 1M tokens.
• Qwen3.7-Plus: Limited-time 20% discount on the alias; no public end date. Standard tier through 256K input tokens. Above 256,000 input tokens, the rate is $0.96 input and $3.84 output per 1M tokens.
• Qwen3.7-Max: Standard international rate through the 1M-input-token tier; no promotional discount assumed.

Cache-hit input is separate from standard input. Alibaba documents typical explicit-cache creation at 125% and explicit hits at 10% of standard input; implicit hits are typically 20%. Modes are mutually exclusive, hits are not guaranteed, and model-specific cache billing governs. Do not stack cache and batch discounts or interpret one cache mode as the other.

The current Singapore / International table lists no free quota for Qwen3.6-Flash. Qwen3.7-Plus and Qwen3.7-Max list a selected-model activation quota of 1 million tokens, valid for 90 days from activation, model release or application approval, whichever is later. This is not a recurring free production allowance and does not establish Qwen Studio plan entitlement.

Cerebras Qwen3.8-27B: limits and trial conditions

The exact Cerebras endpoint is qwen-3.8-27b. Free Trial has 65,536 context tokens and 32,768 maximum output tokens; paid tiers and trial customers have 131,072 context tokens and 40,960 maximum output tokens. Free Trial is limited to 5 requests/minute and 1M tokens/day; Developer lists 300 requests/minute, 150K uncached tokens/minute, temporarily 750K total tokens/minute, and no daily token limit. It accepts text and PNG/JPEG image input and produces text. Free Trial permits 2 images per request, while Developer and Enterprise Shared Inference permit 10; maximum request payload is 10 MiB.

Cerebras states that adding a valid payment method can grant a one-time promotional credit of USD 5, expiring after 30 days. There is no automatic paid enrollment; API and Playground access pause when the credit expires or is exhausted unless PayGo credits are separately purchased. Development/evaluation tier access is not a production entitlement; enterprise capacity and terms require agreement.

Cerebras advertises approximately 1,850 output tokens/s for this route; this is a provider measurement, not an application latency guarantee. Its two current official documents agree on $0.99 input and $1.49 output per million tokens.

Memdoor Qwen3.8-27B team-solve: results and costs

The maintainer reports 49 successful runs on ten problems repeated five times, not a full LiveCodeBench Hard win. Fable 5 was not rerun in this harness. These are recorded OpenRouter backend quotes at the verification date below, not a newly verified current offer, the historical experiment bill or Cerebras prices.

Price verified 2026-11-11 · Official OpenRouter backend listings · Structured prices and launch evidence · How we verify. Listing is not account access proof; routing, cache eligibility and precision matter. Modal precision is unknown. Independent replay remains blocked.

GVS5H coding claims and Cerebras dedicated deployments

• GVS5H: Qwen3.8-27B and Fable coding-cost claims — third-party orchestration, not new model weights or a new hosted rate. The local experiment and the Cerebras route above are different backends.
• Cerebras trillion-parameter deployment and Kimi K3 pricing status — Kimi K3 is listed for enterprise Dedicated Inference, not the shared model catalog. No public Cerebras Kimi K3 token quote is established.

These dated research notes do not renew the unrelated API, Studio or entitlement evidence on this page. Independent replay and dedicated-endpoint blockers remain explicit.

Other recorded Qwen catalog rates

These rows come from the shared API catalog. Their original source-verification dates are shown individually; this editorial review does not refresh them. Overdue rows are recorded rates, not confirmed current offers. Qwen3.8-Flash is excluded while its first-party route and historical announcement require reconciliation.

Qwen Studio is separate from token billing

Qwen Studio chat access is a different product from Alibaba Model Studio API billing. The public Qwen product and chat URLs returned app shells, without meaningful plan-price, quota or model-entitlement evidence. A paid plan, fixed free allowance or zero-dollar production API cannot be verified from these responses.

For local deployment, use the local AI versus API calculator. Downloadable weights are not a free hosted API; hardware, operations and the exact model license still matter.

For coding workloads, see the Qwen Code guide and Qwen versus DeepSeek comparison. Dated launch reports are historical context, not current access evidence.

Frequently asked questions

Official sources and maintenance

• catalog — reviewed semantic source
• pricing — reviewed semantic source
• cache — source changed — review required
• cerebrasModel — reviewed semantic source
• cerebrasPricing — unavailable or unresolved at latest check
• Qwen Studio product plans — unavailable or unresolved at latest check
• Qwen chat plan entitlement — unavailable or unresolved at latest check

Reviewed source projections bind prices, access, limits and billing separately. Source or payload changes freeze verification clocks and remain visible for Data review; Qwen Studio and conflicting routes stay red.