Key takeaways

Two announcements, one signal

On August 3, 2026, Alibaba released Qwen3.8-Max, the new flagship of its AI model family. The spec sheet is massive: 2.4 trillion parameters in an MoE architecture (Mixture of Experts, where only a fraction of the model activates per request), text, image and video input, and a 1 million-token context window. The price is anything but massive: $2 per million input tokens, $6 per million output tokens, and $0.25 for cached input. On Alibaba's published benchmarks, the model edges past Claude Fable 5 on Terminal-Bench 2.1 (86.6 versus 84.6) but trails on coding, with 67.7 against 80.0 on SWE-bench Pro.

The same day, independent evaluation firm Artificial Analysis named DeepSeek V4-Flash the cheapest well-known AI model to run. The rates are dizzying in the other direction: $0.14 per million input tokens, $0.28 per million output tokens, a 75% cut presented as permanent. Running Artificial Analysis's full test suite costs about 3 cents with V4-Flash, versus 86 cents with Kimi K3 (Moonshot AI), $1.86 with GPT-5.6 Sol (OpenAI) and $3.15 with Claude Fable 5 (Anthropic). DeepSeek, which just closed a funding round of more than $7bn, is explicitly targeting high-volume workloads: chatbots, code assistance, back-office automation.

Price per token does not tell you what a task costs

The first useful reflex: never compare models on price per million tokens, but on the cost of the finished task. That is how Artificial Analysis reasons: its measure aggregates the volume of tokens each model actually consumes to complete the same work. A cheap but verbose model, or one that needs three attempts, can end up costing more than a premium model that gets it right the first time. It is the same logic as a contractor's quote: the hourly rate means nothing until you know the number of hours.

The second reflex: read capability scores next to prices. V4-Flash scores 50 out of 100 on Artificial Analysis's Intelligence Index, level with Google's Gemini 3.6 Flash; Kimi K3 reaches 57; the frontier models from OpenAI and Anthropic sit around 66. That 16-point gap is not cosmetic: it separates models that execute a well-scoped instruction from those that carry a chain of reasoning end to end without getting lost. Zack Kass, a former OpenAI go-to-market lead quoted by The Next Web, sums up the dynamic: once models are close enough for a given use, price makes the decision.

The practical consequence fits in one sentence: for every process you want to automate, there is now a minimum level of intelligence that is good enough, and its price just dropped again.

What does this change for a small or mid-sized company?

First, good budget news. The most profitable automations in an SME, sorting and qualifying inbound emails, extracting invoice fields, classifying support tickets, summarizing meeting notes, drafting reply templates, are exactly the kind of well-scoped, high-volume, low-stakes-per-unit tasks that 3-cent models were built for. A process handling 1,000 documents a day that would have cost a few hundred euros per month in inference in 2024 now costs single-digit euros. Model cost has stopped being a reason to postpone an automation project.

Next, a warning. DeepSeek's and Qwen's advertised prices apply to their APIs, which means your data travels to servers governed by Chinese law. For personal use or public data, that is a choice. For customer records, contracts or HR files, it is very hard to defend under GDPR (the EU's General Data Protection Regulation), and the AI Act's timeline, with national sanctions applicable since August 2, 2026, leaves little room for approximation. The dividing line is simple: Chinese APIs for nothing confidential, and other routes for everything else.

That is why the third announcement matters as much as the first two: Alibaba has promised to publish Qwen3.8-Max's weights within a week, and is already releasing Qwen3.8-27B, a compact version, as open weights. An open-weights model can run at a European hosting provider or on your own servers: the Chinese price becomes a global reference point without your data crossing any border. For a European SME, the reasonable setup in 2026 looks like this: a frontier model via API for complex, high-stakes tasks; a small, cheap, open model hosted in Europe when data is sensitive, for volume; and an off-the-shelf tool with no AI at all when the process simply does not need it.

How this connects to my day-to-day work

Matching intelligence level to task is not a theory, it is a trade-off I make on every project. My Bloomberg Dashboard, which analyzes a personal portfolio, deliberately runs on Claude Haiku 4.5, a small model: the tasks are well-scoped, the volume is daily, and a frontier model would add nothing but an invoice. Conversely, the automated newsletter IA Brew and its 93 n8n nodes reserve the most capable model for the writing steps, where quality shows. Sizing the model to the task, rather than defaulting to the biggest one everywhere, has become one of the first cost levers in an AI agent project; the August 3 price war makes that trade-off pay even more.

The takeaway: on August 3, 2026, the price of artificial intelligence dropped again, and the gap between the cheapest model and the most capable one widened again. The useful question for an SME is no longer "how much does AI cost?" but "what level of intelligence does each process require, and where is my data allowed to go?". Companies that can answer process by process will pay cents where others pay licenses.

Frequently asked questions

Can a European SME use DeepSeek or Qwen in production?

Yes, but how you access them changes everything. Calling DeepSeek's or Alibaba's Chinese API means sending your data to servers governed by Chinese law, which is hard to defend under GDPR for customer data. Models released as open weights, however, can run at a European hosting provider or on your own servers: your data never leaves your infrastructure.

Is a less intelligent model enough to automate a business process?

Often, yes. Sorting emails, extracting invoice fields, classifying support tickets or summarizing meeting notes are well-scoped tasks that a model scoring 50 out of 100 on Artificial Analysis's Intelligence Index handles correctly. Frontier models, scoring around 66, remain necessary for multi-step reasoning, complex code and high-stakes decisions.

What do open weights mean for an AI model?

An open-weights model is one whose trained parameters can be downloaded. A company can run it on its own servers or at any hosting provider, without depending on the vendor's API. It is the difference between renting a service and owning the engine: you gain control and confidentiality in exchange for the operating burden.

Will AI model prices keep falling?

The underlying trend points that way: DeepSeek presents its 75% cut as permanent, and each model generation delivers more capability at the same price. There is no guarantee, however, that frontier models will match Chinese pricing; the price gap buys capability, contractual guarantees and data residency.