DeepSeek V4 Pro vs Qwen3.8 Max: Reasoning Specialist vs the Chinese All-Rounder

Ten days apart, the two most-watched models in China went live: Alibaba released Qwen3.8 Max on August 3, 2026 — a 2.4-trillion-parameter sparse-MoE flagship it bills as an all-rounder for agents, office work, and multimodal understanding — and D​eepSeek shipped the official version of DeepSeek V4 Pro on August 12, its reasoning-and-coding specialist with 1.6T total parameters and a price tag roughly one-seventh of Q​wen's on output tokens. The two are aimed at the same developer wallets but disagree about what a flagship is for. Qwen3.8 Max wants to be the one model you route everything through; DeepSeek V4 Pro wants to be the one you escalate the hard stuff to.

Key takeaways

• Qwen3.8 Max is the higher raw-capability model on the Chinese general leaderboards (SuperCLUE #1, July) and the stronger multimodal/enterprise package — video input, built-in tools, structured output, all in one API.

• DeepSeek V4 Pro is dramatically cheaper (~7× on output), already open-weights, and now holds the higher coding/agentic claims — Terminal-Bench 2.1 at 87.9 against Q​wen's vendor-reported 86.6.

• Watch the verbosity cost: independent runs show Qwen3.8 Max averaging ~64 turns and ~$1.14 per agentic task — an effective-cost problem the spec sheet hides.

• The honest headline: Qwen3.8 Max is the "capability war" flagship that matches US closed-model pricing; DeepSeek V4 Pro is the cost-performance rebuttal.

Two philosophies in one spec table

• Total params — Qwen3.8 Max 2.4T (sparse MoE, ~95B active) vs DeepSeek V4 Pro 1.6T (MoE, ~49B active)

• Context / max output — both 1M context; Q​wen tops out around 128–131K output vs D​eepSeek's 384K

• Inputs — Q​wen accepts text, image, and video; DeepSeek V4 Pro accepts text and image, with new native image reasoning in the official build

• Open weights — DeepSeek V4 Pro: released (MIT); Qwen3.8 Max: promised for the week of August 10, the first Max-class Q​wen ever to be open-sourced

• API extras — Q​wen ships built-in tools and structured output out of the box; DeepSeek V4 Pro ships thinking toggles, structured JSON, a Responses API, and Codex compatibility

• Price — Qwen3.8 Max $2.00 / $6.00 per million tokens ($0.25 cached) vs DeepSeek V4 Pro ~$0.42 / $0.87 ($0.0035 cached)

Where Qwen3.8 Max wins

On the generalist axis, Qwen3.8 Max is ahead and the independent data agrees. Artificial Analysis puts it at an Intelligence Index of 58, a Coding Index of 71.8, and an Agentic Index of 58 — the closest any Chinese lab has come to the top of that agentic leaderboard. On SuperCLUE's July Chinese-language ranking it sits #1 at 71.48 while DeepSeek-V4-Pro sits #5 at 64.4. Alibaba's launch demos — a 16-day autonomous coding run, a paper reproduction that beat the original paper's AIME24 result — are the strongest evidence in the whole release cycle that this is a genuinely capable long-horizon agent.

For a builder, the practical wins are integration-shaped: video input, built-in tool calling, structured output without configuration, and an ecosystem (Model Studio, the Q​wen toolchain) that D​eepSeek does not try to match. If the workload is document-heavy, multimodal, or needs an enterprise integration story, Qwen3.8 Max is the model the Chinese market is currently defaulting to.

Where DeepSeek V4 Pro wins

On coding and cost, the official V4 Pro is the rebuttal. D​eepSeek reports the official build at 87.9 on Terminal-Bench 2.1 (up from 72.1 on the preview) and 62.7 on DeepSWE — against Q​wen's vendor-reported 86.6 Terminal-Bench. The preview already held an 80.6 SWE-bench Verified, a 90.1 GPQA Diamond, and a 93.5 LiveCodeBench, the #1 open-model competitive-programming score, and the official build's changes are concentrated exactly in the agentic and reasoning tasks where those numbers live.

Then there is price. At $0.42/$0.87 per million tokens, DeepSeek V4 Pro is about 4.8× cheaper on input and 6.9× cheaper on output than Qwen3.8 Max's $2/$6. D​eepSeek also flagged on August 6 that a price increase is coming, which makes today's rate a window, not a promise — more on that below.

Do the math on a real agentic task

Independent agentic runs are where Q​wen's sticker price stops being the real price. On Artificial Analysis agentic tasks, Qwen3.8 Max averaged ~64 turns per task and cost ~$1.14 per task, versus ~$0.53 for the previous generation — the model is verbose and plans a lot. Run the same task shape on DeepSeek V4 Pro at $0.87 per million output and the per-task cost lands roughly an order of magnitude lower. If your product is a loop that calls the model a hundred times a day per user, that difference is your margin.

Reliability caveats on both sides

Sourcing honesty cuts both ways here. Independent testing flagged Qwen3.8 Max's hallucination rate rising from 23% to 40% and an AA-Omniscience score down ten points — the model got more capable and less trustworthy in the same release. D​eepSeek's preview carried a documented 94% answer-always rate on AA-Omniscience, meaning it tends to produce an answer whether or not it knows one. Both flags matter for production; neither is disqualifying for the workloads these models are aimed at.

Why the open-weights timing changes the price math

Qwen3.8 Max's open-weights release, when it lands, will change this matchup's economics inside a week — self-hosters get a 2.4T flagship, and the API price pressure will be real. This is precisely the scenario where a pass-through pricing model keeps you honest. OrcaRouter passes the provider list price through with zero markup, so the moment Alibaba or D​eepSeek moves a price — up or down — the change is live on one key the same day, with no renegotiation and no second contract to read. You route to whichever of Qwen3.8 Max or DeepSeek V4 Pro your current eval favors, and the pricing stays whatever the vendor actually charges.

Which one for your API builder decision

Pick Qwen3.8 Max when the job needs the whole surface — multimodal input, structured output, built-in tools, Chinese-language quality at the top of the current rankings — and your cost model tolerates verbose agentic runs. Pick DeepSeek V4 Pro when the task is reasoning- or coding-heavy, the token volume is high, open weights matter for compliance or self-hosting, or the budget line is the binding constraint. The cleanest reading of this matchup is the one the two companies themselves are signaling: Qwen3.8 Max is the flagship you measure everything against, and DeepSeek V4 Pro is the one that makes the benchmark look expensive.