Qwen 3.8 API Access: Pricing, Token Plan & Options (2026)

For teams trying to get Qwen 3.8 into an actual application, not just read about benchmarks.

Short version: Qwen 3.8 now has a standalone API. Alibaba launched standard metered access on August 3, 2026, at $2 per million input tokens, $6 per million output, and $0.25 for cached input.

Whether the metered flagship or the self-hosted 27B is the better fit for your workload is the bigger question. We broke down the Max vs 27B decision in full.

What doesn't exist yet: independent benchmarks for the full flagship.

Here's how to build with it today without betting your roadmap on unverified numbers.

TL;DR

Where Qwen 3.8 actually stands

Alibaba announced Qwen 3.8-Max on July 19, 2026 at WAIC in Shanghai: a 2.4 trillion parameter sparse MoE model, multimodal, with roughly 95B active parameters per token, a number Alibaba disclosed at launch. The announcement came with a bold internal claim ("second only to Fable 5") and no published benchmark table. We broke down what's verified and what isn't in our Qwen 3.8 benchmarks post.

That changed on August 3, 2026: Qwen 3.8-Max now has a generally available API with a published rate card. $2 per million input tokens, $6 per million output, $0.25 for cached input. Ordinary metered access is live through Alibaba's platform.

The three ways to touch Qwen 3.8 today

1. The standard API. Live as of August 3, 2026: metered access at $2 per million input tokens, $6 per million output, and $0.25 for cached input. It ships with OpenAI-compatible and Anthropic-compatible API specs like the rest of the Max line, so an existing test harness ports over with minimal changes. This is now the production route.

2. Alibaba's Token Plan. A credit-based monthly subscription, not a metered API. Third-party reporting puts the Personal Edition tiers at roughly $6, $20, and $70 per month (39, 139, and 499 CNY), with an additional 80% off credit consumption between 22:00 and 08:00 UTC+8 (eesel's pricing breakdown has the details). During the preview window, Qwen 3.8-Max reportedly ran at 10% of standard credit consumption; Alibaba was explicit that preview pricing was promotional, so confirm current rates before counting on them. Fine for individual use and evaluation, not for a production SLA.

3. Qoder and QoderWork. Alibaba's coding products carry the model too. If your use case is agentic coding, this is the fastest way to feel the model out.

The fourth option, open weights, has now fully arrived: Qwen3.8-27B is live on Hugging Face under Apache 2.0 (our local setup guide covers it), and a text-only 2.4T Max-class checkpoint, Qwen3.8-2.4T-A95B, landed around August 12-13 under a custom license. The 2.4T is a datacenter deployment (400GB+ even at 1-bit), so for most teams the API remains the way to touch Max-tier capability.

How to run Qwen 3.8 in production

Qwen 3.8-Max now has a rate card, so you can build on it directly through Alibaba's API, or through Yotta AI Gateway if you'd rather keep it behind the same key as the rest of your models. But it still has no verified benchmarks. If that's more single-vendor risk than you want, two realistic paths, depending on how you like to operate.

Managed API today: Qwen 3.7-Max. Still the proven flagship of the family, with published benchmarks (92.4 GPQA Diamond, 80.4 SWE-bench Verified) and real per-token pricing at $1.25 per million input tokens and $3.75 per million output. It's live on the Yotta AI Gateway alongside Claude, DeepSeek, GLM, and the rest of the Qwen line, behind one OpenAI-compatible key. And because Qwen 3.8 speaks the same API specs, a gateway setup means swapping a model string later, not rewriting a client. We compared costs in detail in Qwen 3.7-Max vs Claude Opus 4.6.

Self-hosted today: open-weight Qwen. If your requirement is weights you control, the current open Qwen releases already run well on rented GPUs. Our guide to running Qwen 3.7 in production covers engine choice and GPU sizing, and that playbook now applies directly to the released 27B; our local setup guide covers it.

The wrong move now isn't building on 3.8, it's betting production on a model with no independent benchmarks and a single provider. Point your evals at the new API, and keep production on something verified until independent numbers arrive.

What to watch next

One of the two events we were watching has happened: Qwen3.8-27B's weights landed August 13-14, opening self-hosting and third-party serving for the small model. Still pending: independent benchmark coverage, and whether the text-only 2.4T checkpoint shows up at inference providers, since the flagship's only numbers remain Alibaba's internal evals. We'll update this page as either lands.

FAQ

Is there a Qwen 3.8 API? Yes. Alibaba launched standard API access on August 3, 2026, at published per-token pricing. The Token Plan subscription remains the alternative for individual use.

How much does Qwen 3.8 cost? $2 per million input tokens, $6 per million output, $0.25 for cached input on the standard API.

Is Qwen 3.8 open source? The 27B is, as of August 13-14: Apache 2.0 weights on Hugging Face. A text-only 2.4T Max-class checkpoint, Qwen3.8-2.4T-A95B, is also up under a custom license; the full multimodal Max remains API-only.

Can I download Qwen 3.8 from Hugging Face? The 27B, yes: the official Qwen/Qwen3.8-27B repo is live, and the ecosystem (Ollama, llama.cpp, LM Studio) moved within hours, as it always does for open Qwen releases. The Max-class 2.4T checkpoint, technically yes, but it is text-only and needs cluster-scale hardware; it is not a practical download.

What is Alibaba's Token Plan? A monthly credit subscription for Qwen model access, sold in Lite, Standard, and Pro tiers. It carried Qwen 3.8-Max preview access at a reported 10% credit rate during the preview window; now it's the subscription alternative to the standard API.

Can I get Qwen 3.8 through a gateway or serverless provider? Yes, as of late August: Qwen 3.8-Max and Qwen3.8-27B are live on Yotta AI Gateway behind one OpenAI-compatible API, alongside Claude, DeepSeek, GLM, and the rest of the Qwen line.

Is Qwen 3.8 better than Kimi K3? Early independent signals are mixed, and official benchmarks don't exist yet. We ran the comparison in Qwen 3.8 vs Kimi K3.

What should I build on today if I want Qwen 3.8 later? Anything OpenAI-compatible. The Max line ships OpenAI-spec and Anthropic-spec APIs, so a gateway-based setup today means the 3.8 upgrade is a config change later.

Bottom line

Qwen 3.8 is now a product you can build on, through Alibaba's API at published pricing. What the flagship still isn't: practical to self-host, or independently benchmarked.

Until independent numbers land, an OpenAI-compatible setup keeps the eventual switch to a config change.

Run Qwen 3.8 itself through the Yotta AI Gateway, or check current GPU pricing if you'd rather hold the weights yourself. Full launch coverage in our Qwen 3.8-Max release breakdown.