Standard Compute

Alibaba · updated 2026-08-29

Run Qwen 3.8: providers, pricing and flat-rate access

Alibaba's mid-2026 Qwen generation (3.8-Max and 27B variants) — top-tier open-weight coding performance with aggressive per-token pricing.

Qwen 3.8 runs cheapest per-token through providers like DeepInfra or Alibaba's own API (~$0.4/M input class), most flexibly through OpenRouter, and most predictably through a flat-rate plan like Standard Compute, where Qwen 3.8 serves routine agent steps inside the routed pool at a fixed monthly price.

Your options, honestly

DeepInfra / Alibaba APIPay-per-tokenLowest raw per-token rates; you manage keys, limits and spend alerts yourself.
OpenRouterPay-per-tokenEvery Qwen variant with provider fallback; ~5.5% fee on credits.
Standard Compute$39+/mo flatQwen 3.8 in the routed pool — heavy days cost the same as quiet ones. No model pinning.
Self-host (vLLM)GPU + powerThe 27B runs on a single big consumer GPU; Max-class needs serious hardware.
the flat-rate trade-off

Qwen 3.8's per-token price is genuinely low, so light users should just pay per token. The flat-rate case starts when your agent runs hours daily — context growth makes even cheap tokens add up, and a fixed price removes the variance.

Running it in your agent

All OpenAI-compatible agents (OpenCode, Cline, Aider, Continue, Roo) run Qwen 3.8 with a base-URL swap. On Standard Compute, model 'standardcompute' routes to it automatically for the steps where it's the right tool.

FAQ

Which Qwen 3.8 variant should a coding agent use?
The Max variant for quality-sensitive steps, 27B-class for routine ones. If you don't want to manage that split, routed plans do it per-request.
Is there a subscription for Qwen 3.8 specifically?
Alibaba doesn't sell a Qwen-only coding subscription internationally. The subscription route is a multi-model flat plan that includes it — or self-hosting the open weights.

More: how flat-rate LLM APIs work · per-agent guides for OpenCode, Claude Code and Cline · model comparisons