Qwen 3.8 vs DeepSeek V4: The Value King Keeps Its Crown

For most of 2026, DeepSeek V4 has been the reference point for what AI capability ought to cost: open weights available today, an independently measured Artificial Analysis Intelligence Index of ~52–54, and pricing at $0.435 / $0.87 per million tokens that made every premium flagship look like a luxury purchase.

When Qwen3.8-Max was previewed on July 19 with a 90%-off credits campaign and no published rate, it was impossible to know whether Alibaba intended to compete on price. Now we know. Qwen3.8-Max reached general availability on August 3, 2026 at $2 / $6 per million tokens — aggressive against premium models like Fable 5 and Opus, but roughly 4.6× DeepSeek's input rate and 7× its output rate. This is the one comparison in the series where Qwen's GA pricing is not the good news.

A note for builders — the fastest way to settle this is to run both on your own prompts. OrcaRouter fronts 200+ models behind one OpenAI-compatible endpoint, so you can pit Qwen 3.8 Max against DeepSeek V4 on the same task without wiring up two SDKs.

TL;DR verdict. DeepSeek V4 remains the value king, and GA did not change that — it confirmed it. DeepSeek is ~7× cheaper on output, independently scored at ~52–54, open-weight and downloadable today, fast, and backed by an extremely low cache-hit price. Qwen3.8-Max answers with a much larger scale ambition, a 1M-token flat-rate context, native multimodality DeepSeek lacks, and a self-reported benchmark table headlined by Terminal-Bench 2.1 86.6 and GPQA Diamond 92.6. But it has no independent score, no disclosed active-parameter count, and no released weights. If your reason for looking at Chinese open models is cost efficiency, DeepSeek wins outright. Qwen wins only if you need multimodality or the 1M flat context enough to pay 7× for it.

Key takeaways

• DeepSeek V4 is dramatically cheaper: $0.435 / $0.87 per 1M (AA blended ~$0.18) against Qwen3.8-Max's $2 / $6. That is ~4.6× on input and ~7× on output.

• DeepSeek has been independently scored at an AA Intelligence Index of ~52–54. Qwen3.8-Max remains unscored by every third-party evaluator.

• DeepSeek's weights are downloadable today. Qwen's are promised within the week, with no license or repository yet.

• Qwen's genuine advantages are modality and context: text/image/video input and a 1M-token window at a flat rate with no long-prompt surcharge. DeepSeek V4 is text-focused.

• Qwen's self-reported table is strong — Terminal-Bench 2.1 86.6, GPQA Diamond 92.6, FrontierSWE 73.5 (from a predecessor's 40.7) — but unverified, against DeepSeek's vendor SWE figure of ~80.6% at release.

• The DeepSeek family got cheaper still. DeepSeek V4 Flash shipped on July 31 at $0.15 / $0.29, pushing the floor even lower for coding and agent work.

Accuracy note: all Qwen3.8-Max quality figures are Alibaba-reported and unverified by third parties. DeepSeek V4's index is from Artificial Analysis; its SWE figure is DeepSeek-reported at release. Qwen pricing is the GA rate card and supersedes July's preview discount.

The specs and price, side by side

Here is the hard data, each figure attributed to its source.

• Maker / status — Qwen3.8-Max: Alibaba; GA (August 3, 2026); DeepSeek V4: DeepSeek; released April 2026

• Architecture — Qwen3.8-Max: ~2.4T total, sparse MoE (active params still undisclosed); DeepSeek V4: Open-weight MoE (V4 / V4 Pro lines)

• Context window — Qwen3.8-Max: 1M tokens, flat-rate (983,616 with thinking; max output 131,072); DeepSeek V4: Long-context, text-focused

• Modalities — Qwen3.8-Max: Text, image, video in → text out; DeepSeek V4: Text-focused

• AA Intelligence Index — Qwen3.8-Max: Still unscored; DeepSeek V4: ~52–54 (Artificial Analysis)

• Coding — Qwen3.8-Max: Terminal-Bench 2.1 86.6, FrontierSWE 73.5 (Alibaba-reported); DeepSeek V4: SWE ~80.6% (DeepSeek, at release)

• Pricing (in / out) — Qwen3.8-Max: $2.00 / $6.00 per 1M, flat; cached input $0.25; DeepSeek V4: $0.435 / $0.87 per 1M (AA blended ~$0.18)

• Cache-hit price — Qwen3.8-Max: $0.25 per 1M; DeepSeek V4: Extremely low (DeepSeek)

• Open weights — Qwen3.8-Max: Promised within the week (no license yet); DeepSeek V4: Yes — available now

Two things frame this comparison, and unlike the rest of the series, GA did not shift them in Qwen's favor.

First, the price gap is the story and it is very wide. Qwen's $2 / $6 is a genuinely aggressive number when the comparison is Claude Fable 5 at $10 / $50 or Opus at a blended ~$3.85. Set against DeepSeek V4's $0.435 / $0.87, it is expensive. On a blended basis, Artificial Analysis puts DeepSeek at roughly $0.18 per 1M — an order of magnitude below Qwen. Alibaba priced this model to undercut the Western frontier, not to beat DeepSeek, and the numbers make that intent obvious.

Second, DeepSeek holds the credibility advantage as well as the price advantage, which is unusual. Normally the cheap model is the unproven one. Here DeepSeek V4 has an independent AA index of ~52–54 and weights anyone can download and test, while Qwen3.8-Max has a vendor table and a promise. So a buyer choosing DeepSeek is not trading proof for savings — they get both. Qwen's counter-argument has to be capability and modality, and it has to be large enough to justify a 7× output premium.

What Qwen actually offers for the premium

The honest case for Qwen3.8-Max over DeepSeek V4 rests on three things, none of which is price.

Multimodality. Qwen3.8-Max takes images and video natively and reports OmniDocBench 1.5 92.1 on document parsing and OSWorld-Verified 86.1 on driving a real desktop GUI. DeepSeek V4 is text-focused. If your workload involves screenshots, scanned PDFs, charts, or video, this is not a preference — it is a hard requirement that DeepSeek does not meet, and the price comparison becomes irrelevant.

Flat-rate 1M context. Qwen charges the same $2 / $6 whether the prompt is 5,000 tokens or 900,000. Most providers surcharge long inputs precisely because serving them is expensive. For whole-repository review, contract analysis, or multi-document synthesis, that flat rate can matter more than the per-token rate — though at 7× DeepSeek's output price, the arithmetic still has to work out in your favor.

A higher potential ceiling. This is the speculative one. Qwen's self-reported GPQA Diamond 92.6 and Terminal-Bench 2.1 86.6 are frontier-shaped numbers, well above what a ~52–54 index model would typically produce. If they replicate, Qwen3.8-Max is a meaningfully more capable model and the premium is defensible. The predecessor Qwen3.7-Max scored 46 on the AA index — below DeepSeek V4 — so the entire case depends on a large one-generation jump that no third party has confirmed.

Against those, DeepSeek's advantages are concrete rather than conditional: an audited score, downloadable weights, production-proven speed, and a cache-hit price DeepSeek describes as extremely low. And the floor has dropped further since — DeepSeek V4 Flash arrived on July 31 at $0.15 / $0.29 per 1M, tuned for coding and agentic work, which makes the DeepSeek family cheaper again relative to Qwen than it was a week ago.

Cost in practice: where 7× actually lands

Take a code-analysis pipeline: 120,000 tokens of context, 8,000 tokens of output per call.

• DeepSeek V4: 0.12M × $0.435 = $0.052, plus 0.008M × $0.87 = $0.007. ≈ $0.059 per call.

• Qwen3.8-Max: 0.12M × $2 = $0.24, plus 0.008M × $6 = $0.048. ≈ $0.29 per call.

• At 5,000 calls/day: DeepSeek ≈ $295/day; Qwen ≈ $1,440/day. Over a month: roughly $8,800 versus $43,200.

• Even with Qwen's cached input at $0.25/1M reducing its call to ~$0.08, DeepSeek's own extremely low cache-hit pricing keeps it ahead.

A $34,000/month difference on one pipeline is the kind of number that decides architectures. To justify Qwen here you would need it to be materially better at the task — not marginally — or to need multimodality DeepSeek cannot provide.

The one place the arithmetic genuinely turns is very long contexts. If your prompts routinely run into the hundreds of thousands of tokens and your alternative provider surcharges them, Qwen's flat rate can close part of the gap. But it starts 7× behind on output, so it needs a lot of help.

Openness: both open in principle, one open in fact

Both of these are open-weight lines in intent, which makes the timing difference the whole story. DeepSeek V4's weights are out — you can download, quantize, fine-tune, and serve them today, and the community has been doing exactly that for months. Qwen3.8-Max's weights are promised within the week, with no model card, license, or repository yet. Until a license exists, commercial usability is formally undetermined.

And at flagship scale, Qwen's openness is largely theoretical anyway: 2.4T is roughly 1.2TB at 4-bit against about 141GB per H200, so eight or more top-end accelerators before you serve a token — with an undisclosed active-parameter count that makes throughput unmodellable. DeepSeek V4 is far more tractable to self-host, which is a large part of why it became the default open model.

The concurrently announced Qwen3.8-27B is Alibaba's real answer here. Also going open-weights and reportedly running in about 17GB of VRAM, it competes on DeepSeek's actual turf — deployable open models — rather than on flagship scale. That is the Qwen release worth watching if self-hosting is your goal.

FAQ

Is Qwen 3.8 better than DeepSeek V4?

Unproven, and it costs far more. Qwen3.8-Max's self-reported numbers (GPQA Diamond 92.6, Terminal-Bench 2.1 86.6) suggest a higher ceiling than DeepSeek V4's audited ~52–54 index, but no third party has verified them, and the predecessor Qwen3.7-Max scored 46 — below DeepSeek. DeepSeek is proven, open, and ~7× cheaper on output.

Which is cheaper?

DeepSeek V4, decisively. It lists $0.435 / $0.87 per 1M (AA blended ~$0.18) against Qwen3.8-Max's $2 / $6 — roughly 4.6× cheaper on input and 7× on output. DeepSeek V4 Flash, released July 31, goes lower still at $0.15 / $0.29.

Did Qwen 3.8 Max get cheaper at GA?

It got a real, budgetable price for the first time — $2 / $6 flat across the 1M-token context, replacing July's temporary 90%-off credits campaign. That is aggressive against premium Western flagships but well above DeepSeek's rates.

What does Qwen 3.8 offer that DeepSeek V4 does not?

Native multimodality — text, image, and video input, with a self-reported OmniDocBench 92.1 for document parsing and OSWorld-Verified 86.1 for GUI operation — plus a 1M-token context billed at a flat rate with no long-prompt surcharge. DeepSeek V4 is text-focused, so for image, video, or screenshot workloads it is not a substitute at any price.

Which has better coding performance?

Hard to compare directly, because the reported benchmarks differ. DeepSeek V4 reports SWE ~80.6% at release; Alibaba reports FrontierSWE 73.5 and Terminal-Bench 2.1 86.6 for Qwen. Those are different tests with different difficulty curves, and both are vendor-reported, so neither settles it. DeepSeek's independent index score is the more trustworthy signal available.

Can I self-host both?

DeepSeek V4 yes — weights are out and widely deployed. Qwen3.8-Max not yet, and not practically even after release: ~1.2TB in 4-bit needs eight or more H200-class GPUs, with no license published. The concurrently announced Qwen3.8-27B (reportedly ~17GB of VRAM) is the realistic Qwen option for self-hosting.

Which should I use today?

DeepSeek V4 for almost any cost-sensitive text or coding workload — it is proven, open, fast, and an order of magnitude cheaper on a blended basis. Choose Qwen3.8-Max when you need multimodal input or a flat-rate million-token context, and you have verified on your own tasks that the capability difference justifies roughly 7× the output cost.

Bottom line

Qwen 3.8 vs DeepSeek V4 is the comparison where Qwen's GA pricing works against it. Every other article in this series can point to $2 / $6 as the aggressive number that reprices the frontier. Here it is the expensive option — 4.6× DeepSeek's input rate and 7× its output rate — against a model that is independently scored, open today, fast, and cheap enough that most teams never think about its bill.

Qwen3.8-Max's real answers are modality and context: images and video that DeepSeek does not handle, and a million-token window billed flat. Those are genuine differentiators, and for document or screenshot pipelines they are decisive regardless of price. But the raw value crown was never in play, and DeepSeek V4 Flash's July 31 arrival at $0.15 / $0.29 pushed the floor lower still. Watch two events: the first independent Intelligence Index score for Qwen3.8-Max — which would tell us whether the premium buys real capability — and the Qwen3.8-27B weights, which is where Alibaba will actually compete with DeepSeek. Until then, DeepSeek keeps the crown, and Qwen is the model you add for the modalities DeepSeek cannot reach.