The price of a model doesn’t reflect what it costs a fashion house

The arrival of Kimi K3, the drop in OpenAI’s pricing, and the legal settlement reached by Anthropic are shifting the competition to a new playing field. For fashion houses, choosing an artificial intelligence now involves data sovereignty, the value of archives, and the authenticity of the creative signature.

When Intelligence Becomes a Raw Material

On July 16, 2026, Moonshot AI unveiled Kimi K3, a Chinese open-weight model with 2,800 billion parameters. Its “Mixture-of-Experts” architecture activates only a fraction of these parameters for each query: 16 experts out of 896. This technical efficiency allows it to deliver performance comparable to the most advanced American models, while offering significantly lower usage costs.

Kimi K3 notably took the lead in the Frontend Code Arena, which focuses on interface generation. Broader comparisons remain mixed: it does not outperform all models on every test. But that was enough to reignite concerns that the markets had already been feeling since DeepSeek. On July 17, the Nasdaq fell 1.4%, in a move also fueled by questions about tech valuations and the profitability of investments in AI.

A few days later, OpenAI responded with price cuts. On July 30, the company announced an 80% price reduction for GPT-5.6 Luna and a 20% reduction for GPT-5.6 Terra. Uses that were previously limited to a few experimental projects can now be incorporated into the regular operating costs of a marketing department, a studio, or a customer service department.

The luxury industry has never chosen its materials based solely on price per kilogram. It would be strange if it chose its intelligence based on the price of a million tokens.

Origin is becoming just as important as performance

The price listed by Kimi K3 is attractive. Its origin raises more questions.

U.S. officials accuse Moonshot AI of using a distillation technique to replicate certain capabilities of models developed by Anthropic. Moonshot disputes these allegations, which have not been publicly substantiated. Nevertheless, the debate is revealing: so-called “open” models sometimes disclose their parameters, but rarely the complete lineage of their data and training.

For a fashion house, this lack of transparency is not an abstract concept. A model may receive collection archives, previously unpublished images, development documents, relational data, or materials related to an upcoming campaign. The decision is therefore no longer the sole responsibility of the IT department.

Before any deployment, four questions should be raised with management: Where is the data processed? Under which jurisdiction? Can it be retained or used to improve the model? What options are available for reverting to the previous system if the company switches providers?

Sovereignty does not necessarily mean rejecting foreign technologies. It begins with knowing exactly what we are entrusting to them.

Three thousand dollars isn’t the price of a book

The other development of the week comes from the U.S. court system. On July 20, a federal judge definitively approved the $1.5 billion settlement reached between Anthropic and authors whose works had been obtained from pirated libraries.

The settlement covers approximately 500,000 eligible titles out of more than seven million downloaded copies. Rights holders are to receive at least $3,000 per title, before any distribution among multiple rights holders and the deduction of certain expenses.

This figure does not represent a licensing fee. It does not mean that a fashion house’s catalog, photograph, or archive is now “worth” $3,000. It reflects the negotiated settlement amount to resolve a dispute related to an illegal acquisition. This distinction is essential.

The case also distinguishes between two practices that are often confused. The court had upheld, under U.S. fair use law, the use of legally acquired books for training purposes. However, it had refused to protect the creation of a library using pirated copies. The physical copies purchased by Anthropic had been cut up, scanned, and then destroyed; the pirated files came from a different source.

For the institutions, the lesson lies less in the amount of the settlement than in the transformation of their heritage into a financial risk. An uncataloged archival collection, without a licensing policy or rules governing its use by AI, is no longer merely a dormant asset. It becomes a platform for exposure.

The AI Act requires traceability, not a blanket admission

In addition to this issue of origin, there is now the issue of transparency. The requirements set forth in Section 50 of the AI Act will take effect on August 2, 2026.

The rule is more specific than simply stating that “all AI-generated content must be labeled as such.” In particular, providers of generative systems must make their output detectable in a machine-readable format. Companies that deploy these systems must, in certain situations, flag deepfakes and generated text to inform the public about matters of general interest. Exceptions apply to artistic, creative, or satirical works.

For a company, it would be a mistake to wait for a perfect set of guidelines before taking action. The immediate challenge is to establish internal traceability: the tools used, the origin of the elements, the level of human intervention, the rights associated with the source data, and the final destination of the content.

Transparency should not take the form of an awkward note added as an afterthought. When done right, it can build on a standard that the luxury industry already upholds: documenting origin, craftsmanship, and accountability.

Govern Before Generalizing

Technology is finally becoming less visible just as it gains greater autonomy. Claude Opus 5 focuses on agents capable of performing long-running tasks. At the same time, the expanded partnership between Anthropic and Cognizant shows that these models are entering organizations through consulting firms and integrators just as much as through senior management.

OpenAI is observing the same shift across different professions. In an analysis of more than 800,000 messages from U.S. workers, 43.5% of the uses associated with a particular profession involved tasks traditionally associated with another profession. AI is therefore not merely replacing certain tasks; it is quietly redrawing the boundaries between job functions.

A communications team that analyzes data, a studio that produces its own prototypes, or a customer service department that writes campaign content can become more autonomous. They may also inadvertently bypass legal, creative, or strategic approvals.

The main issue is no longer adoption. It’s arbitration.

A company should know which processes it wants to accelerate, which actions must remain under human control, and which data must never cross certain boundaries. The best model will not necessarily be the most powerful or the least expensive. It will be the one whose origin the company can explain, whose use it can govern, and whose presence it can embrace as part of its own identity.

Sources

European Commission, “Guidelines on Transparency Obligations for Providers and Deployers of AI Systems,” July 20, 2026.

OpenAI, “Advancing the Price-Performance Frontier with GPT-5.6,” July 30, 2026; “How AI Is Expanding What People Do at Work,” July 27, 2026.

Anthropic, “Introducing Claude Opus 5,” July 24, 2026; “Cognizant and Anthropic Expand Their Partnership,” July 27, 2026.

Reuters, “U.S. Judge Approves Anthropic’s $1.5 Billion Settlement,” July 20, 2026; coverage of the Kimi K3 launch and market movements, July 17, 2026.

Associated Press, “Judge Approves $1.5 Billion Copyright Settlement Between Anthropic and Authors,” July 21, 2026.

Cette publication est également disponible en :