DeepSeek Is Raising Prices. The Cheap Tier Was Capacity, Not a Price.

Bloomberg reported on 6 August 2026 that DeepSeek intended to raise the prices of its API services significantly, and the confirmation followed within days, with the company urging users to plan accordingly and declining to name a figure. To understand why that sentence matters more than the usual pricing update, look at the numbers it is moving away from. DeepSeek V4 Flash has been serving at roughly 0.14 US dollars per million input tokens and 0.28 US dollars per million output tokens, which is not merely competitive, it is the anchor that an entire class of cost assumptions has been resting on. Teams building summarisers, classifiers, enrichment pipelines and background agents have spent the last year sizing their unit economics against a number in that range, and a great many of those business cases have no headroom in them at all.

The interesting part is the reason, because it is not the ordinary story of a vendor discovering it undercharged. A week before the warning, DeepSeek released V4 Flash 0731, a lightweight model at roughly 284 billion parameters that Epoch AI currently ranks as the second most capable open weight model available, behind only Kimi K3 from Moonshot AI. On 1 August, OpenCode reported that V4 Flash had processed 8 trillion tokens in a single day, 5 trillion of them on free usage and 3 trillion on paid traffic. Several analysts read the price signal the obvious way: when the binding constraint is GPU capacity rather than margin, raising the price is the fastest available throttle, because it removes the incentive to spam a serving fleet that cannot be expanded on the timescale demand is growing. That reading is an interpretation rather than a company statement, but it fits the sequence, and it carries a harder implication than a simple rate rise. A price set by scarcity of hardware is not a price you can plan against, because it has no floor and no ceiling that anyone outside the vendor can compute.

Set the DeepSeek move next to the rest of the August board and the picture gets clearer. OpenAI cut GPT-5.6 Luna by around 80 per cent to 0.20 US dollars per million input tokens. Anthropic has been running Claude Sonnet 5 at 2 and 10 US dollars per million through 31 August before it reverts to 3 and 15. One vendor cut hard, one vendor let an introductory rate expire on a published date, and one vendor warned of an unspecified rise, all inside five weeks. Those are not three data points about a market finding its level. They are three separate confirmations that per token pricing in 2026 is a commercial lever rather than a reflection of the cost of serving, and that any figure sitting in a financial model right now has roughly the durability of a promotional code.

The rational response to volatile pricing, and the one plenty of engineering teams reached for during the past year, is to abstract the provider away and route each request to whatever endpoint is cheapest for the job. As an engineering decision it is defensible. As a governance posture it is quietly remarkable, because it means the identity of the supplier processing your production data can change without a ticket, a review, an approval or a record. ISO 27001 Annex A control 8.32 expects changes to information processing facilities to be planned, tested and approved. Annex A 8.9 expects a configuration baseline that reflects what is actually deployed, and a routing table with a cost heuristic in it is a baseline that describes an intention rather than a state. If you report against SOC 2, criterion CC8.1 is the change management criterion, and an auditor is entitled to ask which supplier processed customer data last Tuesday. A router that optimises on price cannot answer that question after the fact unless somebody deliberately built the logging to answer it, and in most implementations nobody did.

For DeepSeek specifically there is a second layer that price sensitivity tends to bury. Inference against the hosted DeepSeek API has routed through data centres in mainland China, including for users in jurisdictions with strict residency expectations, and the regulatory response has been among the sharpest that any AI vendor has attracted. The Italian data protection authority, the Garante, moved to block the service, investigations opened across a spread of European jurisdictions, the European Data Protection Board stood up a dedicated AI enforcement effort, and government device bans spread from Washington to Canberra, with the No DeepSeek on Government Devices Act progressing in the United States Congress. DeepSeek did respond on the transparency point, publishing a revised privacy policy in February 2025 with a jurisdiction specific supplemental clause for the European Economic Area, Switzerland and the United Kingdom that maps purposes to legal bases and states plainly that data is stored in China. That disclosure resolves the question of what happens. It does not resolve whether your organisation is permitted to let it happen.

The framework mapping is unambiguous once you look at it as a transfer rather than as a line item. ISO 27001 Annex A 5.19 and 5.21 cover supplier relationships and ICT supply chain security, and an inference provider receiving your prompts is a processor in the plainest reading, regardless of how little the invoice says. Annex A 5.34 covers privacy and protection of personally identifiable information, which is engaged the moment a prompt carries customer content. Under the GDPR, sending personal data to a provider hosting in China is a Chapter V transfer requiring a lawful mechanism and a transfer impact assessment, and the assessment is the part teams skip because the integration took twenty minutes. For Australian organisations, Australian Privacy Principle 8 applies to cross border disclosure and carries accountability for the acts of the overseas recipient, which means the cheap endpoint does not become the problem of someone else once the request leaves your network. Under ISO 42001, the hosted model is a component of an AI system in your inventory, subject to supplier due diligence and to lifecycle control over changes, and swapping the component because the price moved is exactly the change the standard expects you to manage.

None of this will appear in your compliance dashboard, and the reason is structural rather than a missing connector. Vanta, Drata, Secureframe, Sprinto and Thoropass build continuous monitoring on systems that emit records a connector can subscribe to, the identity provider, MDM, cloud accounts, code repositories and the HR system, and that coverage is genuinely strong for what it covers. A model routing rule inside an application configuration emits nothing. There is no tenant, no OAuth consent, no SCIM record, and the billing line, if there is one at all, is a credit card charge from an API vendor that looks identical whether it processed one prompt or ten million. The control that catches this is one you author yourself, with a named owner, a written policy on which providers are permitted for which data classes, and evidence drawn from code review and request logging rather than from an integration screen.

The practical work is a week and starts with a distinction that gets collapsed far too often. DeepSeek the hosted API and DeepSeek the open weight model are different suppliers with different risk profiles that happen to share a brand. Running the open weights on infrastructure you control removes the cross border flow entirely, because there is no DeepSeek server in the path, and it correspondingly hands you the evaluation, patching, safety testing and capacity planning that the hosted service was performing. That is a real trade, and for a document classification pipeline touching customer records it is frequently the right one. Beyond that, do four things. Enumerate every model endpoint in production, with an owner, a data class and a hosting jurisdiction against each, in the same list you should already hold for model identifiers. Write down which providers are approved for which data classes and make the router obey that list before it obeys the price. Log which provider served each request, so the change management question has an answer. Then rerun the financial model at twice the current rate for every workload that depends on the cheap tier, because the vendor has told you in advance that the number is moving and has pointedly not told you how far.

The wider lesson is the same one this industry has now relearned with cloud, with SaaS, with containers, with open source packages, with MCP servers and with model deprecation, and the shape never varies. A genuinely useful capability arrives, adoption outruns the governance around it, and the correction lands as an incident rather than as a decision. What is different about cheap inference is that the correction has been announced ahead of time, in public, by the vendor. The organisations that handle the next few months well will be the ones that treat a price warning as what it actually is, notice that a supplier holding your production data has told you its economics are not stable, and use the interval to find out where that supplier sits in their stack before the invoice, or a regulator, tells them.