Sakana Namazu vs Qwen 3.8: One Is Live, the Other Is Announced

Sakana Namazu and Qwen 3.8 arrived in the same seven days of 2026, and both carry the weight of what their home country hopes a national LLM will be. Sakana Namazu is Japan's: a commercial API live since August 2, built on the open-weight Kimi K2.6 and tuned for Japanese business language, at $0.95 / $4.00 per million tokens. Qwen 3.8 is China's: Alibaba's next flagship, announced July 19, with a preview already running on Alibaba's own platforms and the full release and open weights still pending. Only one of them is callable today.

Two flagships, a week apart

The timeline is the story. July 19: Alibaba announces Qwen 3.8, a reported 2.4-trillion-parameter mixture-of-experts flagship with open weights promised and no license or date. August 2: Sakana opens the Sakana Namazu API, a finished product with pricing, tools, and an enterprise console. August 3: Qwen 3.8-Max-Preview appears on Alibaba's Token Plan, Qoder, and QoderWork at a reported $2 / $6 per million tokens — but general availability, open weights, and any routing-platform presence are still ahead. It is a race where one runner has already crossed the line and the other is still being announced.

What you can actually build today

With Sakana Namazu, the answer is everything: an OpenAI-compatible endpoint, web search and code execution as first-class tools, function calling, and Sakana's demonstrated agent loop (plan → search → cross-check → produce). With Qwen 3.8, the answer is: wait. The preview runs on Alibaba's own products, and the model ID is wired for routing — OrcaRouter's catalogue lists qwen/qwen3.8 with the note that calls return model_not_yet_available until launch and then route automatically at provider cost — but "the day it goes live" is not a shipping date. For a team choosing an architecture this quarter, live beats announced.

The spec comparison

• Status — Sakana Namazu live Aug 2, 2026 vs Qwen 3.8 announced July 19, preview-only, GA pending

• Price — Namazu $0.95 / $4.00 per 1M tokens vs Qwen 3.8 $2 / $6 per 1M tokens (reported, unconfirmed)

• Scale — Namazu inherits Kimi K2.6's 1T / 32B-active MoE vs Qwen 3.8 reported 2.4T total / ~95B active

• Context — Namazu not disclosed vs Qwen 3.8 ~1M tokens (reported, unconfirmed)

• Open weights — Namazu closed vs Qwen 3.8 promised, not yet released

• Verification — both vendor-reported only; neither has independent scores

The sovereignty question

The subtext of both models is national. Sakana Namazu's headline benchmark is FairPoliticsQA, its internal test of political neutrality — a 34.10% to 56.30% jump over the base model that is explicitly about not answering like a model trained on another country's internet. It is built on a Chinese base (Kimi K2.6) but tuned to answer like a Japanese business counterpart. Qwen 3.8, meanwhile, is the flagship of China's open-weights ecosystem, with vendor claims of frontier-class coding (GPQA Diamond 92.6, Terminal-Bench 2.1 86.6, SWE-bench Pro 67.7 — all self-reported) and content guardrails shaped by Alibaba's regulatory context. For an enterprise choosing a foundation, "whose framing does this model carry?" is now a procurement question, not a political one.

What the price difference means

Even taking Qwen 3.8's reported $2 / $6 at face value, Sakana Namazu is the cheaper call at $0.95 / $4.00 — and it is live. The interesting wrinkle is that Namazu's price is exactly the base model Kimi K2.6's price: Sakana charges no premium for its post-training, which makes the Japanese-tuned version the default pick over calling the raw base. Qwen 3.8's reported price is a preview-era number; the durable rate, like the open-weights license, is unannounced. On cost certainty alone, Sakana Namazu is the safer line item.

The decision

Choose Sakana Namazu if you need a working Japanese-business model this month, if your workload is keigo-sensitive or politically neutral in the Japanese sense, or if a hosted API with tools beats a roadmap. Choose Qwen 3.8 if you are building for scale where a reported 2.4T-parameter frontier matters more than availability, if you want open weights you can self-host (when they actually drop), or if your stack is Chinese-market-facing. And if you are hedging — the sane posture when one side is live and the other is promised — note the asymmetry is already wired: Qwen 3.8 will route through OrcaRouter the day it launches at provider cost with no markup and no code change, so you can adopt Sakana Namazu today and add Qwen 3.8 the morning it lands, from the same key.