China's AI Price War: How Alibaba's Qwen3.8-Max and DeepSeek V4-Flash Are Reshaping Inference Costs

Alibaba has shipped Qwen3.8-Max, the largest model it has ever put out, at roughly the same moment that DeepSeek's V4-Flash started picking up attention for undercutting a long list of rival systems on inference price. Read together, the two releases say something about where Chinese AI development is heading: scale is still a headline, but cost is increasingly the pitch.

Key takeaways:

• Alibaba's Qwen3.8-Max is its biggest model yet at 2.4 trillion parameters, but only about 95 billion fire on any given request.
• DeepSeek V4-Flash is competing on a different axis entirely: it lists at $0.14 per million input tokens and $0.28 per million output tokens.
• Sticker price per token is a weak predictor of what a job actually costs. Output volume and the number of round trips matter just as much.
• All three major Chinese labs in this story publish open weights, which is a deployment choice that Western frontier labs have largely declined to make.

Two Different Bets on What Wins

The two companies are not making the same bet. Alibaba went big and is arguing that a very large model can still be economical to serve. DeepSeek went small and is arguing that most workloads never needed the big model in the first place.

Inside Qwen3.8-Max: Big Model, Small Slice

Qwen3.8-Max carries 2.4 trillion parameters, but that figure alone is misleading. The model uses a mixture-of-experts (MoE) design, in which a routing layer wakes up only a portion of the network for each incoming request. Alibaba puts the active count at roughly 95 billion parameters at any one moment, which trims both serving cost and latency compared with lighting up the whole thing.

Capability-wise, the model handles text, images, and video, and accepts context windows stretching to one million tokens. Alibaba also pointed to a demonstration in which Qwen3.8-Max worked through a software engineering project autonomously across a 16-day stretch.

On sheer size it lands just under Kimi K3, the model Moonshot AI released in July, which totals 2.8 trillion parameters with roughly 104 billion active. The two are also going head-to-head on price:

• Qwen3.8-Max — Input $2.00 / 1M tokens, Output $6.00 / 1M tokens, 2.4T total params, ~95B active params
• Kimi K3 — Input $3.00 / 1M tokens, Output $15.00 / 1M tokens, 2.8T total params, ~104B active params
• DeepSeek V4-Flash — Input $0.14 / 1M tokens, Output $0.28 / 1M tokens, 284B total params, 13B active params

After launch, Qwen3.8-Max climbed to the top of the Chinese text-model rankings on the crowdsourced comparison site Arena.AI, though it still sits behind several Anthropic entries in the overall standings. On Arena.AI's board for models that interpret images and other visual input, it finished second worldwide, edged out by a Claude Fable 5 variant.

DeepSeek's Alternative Route: Compete on the Bill

DeepSeek did not chase the parameter counts coming out of Alibaba and Moonshot AI. It went the other way and priced V4-Flash beneath a broad set of widely deployed systems.

Artificial Analysis lists V4-Flash at $0.14 per million input tokens and $0.28 per million output tokens, with a one-million-token context window, 284 billion total parameters, and 13 billion active during inference — a far sparser ratio than either of its larger domestic rivals.

The number that stands out most is cached input. For the Max Effort configuration of V4-Flash, Artificial Analysis records cache-hit pricing at $0.003 per million tokens, about 98% below the standard input rate. Cached input refers to context the system has already processed and can reuse on later calls, which is exactly the pattern that agent loops, retrieval pipelines, and chat products fall into naturally when they resend the same system prompt or document set.

Those rates carried straight through into benchmark testing. Reuters reported Artificial Analysis figures putting V4-Flash's average cost at three cents per test. The same measurement came in at 86 cents for Kimi K3, $1.86 for OpenAI's GPT-5.6 Sol, and $3.15 for Anthropic's Claude Fable 5.

On capability, the Max Effort reasoning variant of V4-Flash scored 40 on the Artificial Analysis Intelligence Index, and the firm clocked throughput at roughly 118 output tokens per second.

Why Token Price Is Not the Same as Task Price

Here is the part most pricing comparisons skip. A per-token rate is an input to the final bill, not the bill itself. What you actually pay depends on architecture, active parameter count, how many tokens the model burns getting to an answer, and how many separate calls the job takes to finish.

Kimi K3 illustrates the gap well. Its list price is $3 per million input tokens and $15 per million output tokens, with cached input at $0.30 per million. But on AA-Briefcase, the Artificial Analysis benchmark built around agentic knowledge work, Kimi K3 averaged $10.57 per task. Getting there took roughly 120,000 output tokens and an average of 83 turns per task.

Artificial Analysis attributed that figure to the combination of token pricing, output volume, and interaction count. A model that talks more, or that needs more back-and-forth to converge, will quietly outspend a nominally pricier model that gets there in fewer moves.

None of this makes Kimi K3 a poor performer. It posted the second-highest overall result on AA-Briefcase at the time of testing, trailing only Claude Fable 5, and scored 57 on the wider Intelligence Index. The point is narrower: cost-per-task and cost-per-token are different metrics, and only one of them is on the pricing page.

For anyone budgeting a deployment, the practical implication is straightforward. Benchmark your own workload. Two models with the same headline rate can produce meaningfully different invoices depending on how verbose they are and how many turns your pipeline requires.

Open Weights Change the Cost Equation Again

Pricing is only one lever Chinese developers are pulling. Distribution is the other.

Alibaba, DeepSeek, and Moonshot AI have all continued shipping open-weight releases alongside their hosted APIs. Artificial Analysis lists DeepSeek V4-Flash as an open-weight model under an MIT licence with weights on Hugging Face. Kimi K3 is likewise available as open weights, under Moonshot AI's own licence.

That matters because open weights break the dependency on a single hosted endpoint. Teams can run the model on their own hardware, or route it through whichever third-party inference provider fits their budget and latency profile. Infrastructure costs do not vanish — they shift onto whoever provisions the GPUs — but the access path is no longer controlled by one vendor's price list.

It is a clear divergence from the approach at OpenAI, Anthropic, and Google, whose flagship models remain closed.

Lian Jye Su, chief analyst at Omdia, framed the demand side of this plainly: a large share of business workflows simply do not require the highest-scoring model on the leaderboard. What those workflows need, he said, is something capable enough, affordable, transparent, and accessible — and open-weight releases fit that description.

What This Means If You're Choosing a Model

The competitive story here is no longer just about who posts the biggest parameter count. Three variables are now in play simultaneously: raw capability, cost per completed task, and whether you can host the thing yourself.

For high-volume, repetitive work — support triage, classification, RAG over a stable corpus — the caching economics on something like V4-Flash are hard to argue with. For long agentic runs where quality compounds across dozens of turns, a stronger and pricier model may finish in fewer steps and cost less overall. And for regulated or latency-sensitive environments, open weights may matter more than either number.

Measure your own workload before you commit to a line item.

FAQ

How many parameters does Qwen3.8-Max have?

It has 2.4 trillion total parameters, with roughly 95 billion active per request thanks to its mixture-of-experts architecture.

How much does DeepSeek V4-Flash cost?

Artificial Analysis lists it at $0.14 per million input tokens and $0.28 per million output tokens. Cache-hit input on the Max Effort version drops to about $0.003 per million tokens.

Is Qwen3.8-Max cheaper than Kimi K3?

On list price, yes. Qwen3.8-Max is $2 per million input tokens and $6 per million output, against $3 and $15 for Kimi K3. Actual cost per completed task depends on output volume and turn count.

What is mixture-of-experts and why does it lower cost?

MoE splits a model into specialised sub-networks and activates only a few per token. You get the knowledge of a large model while paying compute closer to that of a much smaller one.

Are these Chinese models open source?

DeepSeek V4-Flash is available as open weights under an MIT licence via Hugging Face. Kimi K3 is open-weight under Moonshot AI's own licence. Alibaba has committed to publishing Qwen3.8-Max weights as well.

Why is cost per task different from cost per token?

Because a model's total bill reflects how many tokens it generates and how many separate calls a job needs. Kimi K3 averaged $10.57 per task on AA-Briefcase despite a mid-range token rate, using around 120,000 output tokens across 83 turns.