DeepSeek Raises API Prices by up to 1,100%: The Case for Owning Your AI

On 13 August 2026, DeepSeek confirmed it would raise API prices for its V4-Pro and V4-Flash models, with increases running from 50 percent at the mild end to 1,100 percent at the sharp end depending on the model, the token type and the time of day. The new rates take effect on 17 August, four days after the announcement. The company also introduced peak and off-peak billing, so the same request now costs a different amount depending on when your systems happen to run it.

The detail matters more than the headline. For V4-Pro during peak hours, uncached input pricing moves from 3 yuan to 9 yuan per million tokens, and output pricing goes from 6 yuan to 27 yuan per million tokens. Off-peak rates land at 4.5 yuan for input and 13.5 yuan for output. This is the provider that built its entire market position on being the cheapest credible option available. Any business that budgeted a 2027 AI programme on last month's DeepSeek rate card has just watched that budget stop being real, with four days' notice and no negotiation.

Aphelion AI is a private enterprise AI platform built to run capable models inside infrastructure you own or exclusively control, on a flat per user fee rather than a meter, so that a supplier's pricing decision is not also your pricing decision. That is not a reaction to this week's news. It is the reason the platform was built the way it was, and this week is simply the clearest illustration yet of what happens when a business rents the thing it depends on.

What Actually Changed This Week

Strip away the framing and three things happened at once, each of which lands differently on an operational deployment:

• A step change in unit price. Output tokens on V4-Pro at peak hours went up more than fourfold. Output is where agentic and reasoning workloads spend most of their budget, so the headline range understates the impact for exactly the use cases enterprises are moving into.
• Time of day pricing. Peak windows are set daily in Beijing time, which for a European or UK business means the expensive hours are decided by a timezone your operations do not run in. A batch job that was cheap on Friday can be expensive on Monday without anyone changing a line of code.
• Four days of notice. There is no meaningful window in which to re-architect, re-benchmark an alternative model, or renegotiate. The change is announced and then it is live.

None of this makes DeepSeek a bad actor. Inference genuinely costs money on every single call, and a provider that priced below cost to win the market was always going to correct eventually. The point is not that the supplier behaved badly. The point is that the supplier had the unilateral ability to reset your cost base, and used it.

This Is Not One Vendor's Decision

DeepSeek is the sharpest example, not an isolated one. The direction of travel across the market has been clear for most of 2026. OpenAI's flagship developer pricing rose substantially between GPT-5.1 and GPT-5.2, and Anthropic moved its enterprise tier from fixed to dynamic usage based billing in April, a shift analysts expected to double or triple costs for heavy users. GitHub tightened compute allowances on individual Copilot plans and dropped access to certain premium models entirely. Industry cost analysts are advising enterprises to budget for 30 to 50 percent API price rises over the next eighteen months as providers move toward sustainable unit economics.

What makes this genuinely dangerous is the second half of the story: switching is much harder than buyers think. In a survey of more than five hundred US executives with active AI vendor contracts, close to 90 percent believed they could move to a different provider within four weeks, and 41 percent thought they could do it inside a week. Anyone who has actually migrated a production agent estate knows those numbers are fantasy. Prompts are tuned to a model's quirks, evaluation suites are calibrated against its outputs, and tool calling behaviour differs enough between families that a swap is a rebuild, not a config change.

That combination, rising prices plus real lock-in, is what turns a pricing announcement into a business risk. And it explains the more counterintuitive finding in this year's cost research: per token prices have fallen dramatically over two years while total enterprise AI spend has risen sharply, because consumption grew faster than unit prices fell. The FinOps Foundation found that most enterprises overshot their original AI cost projections. Cheaper tokens did not produce cheaper AI.

The Cost You Cannot Put on an Invoice

Price is the visible half of the argument. The other half is what leaves your building every time you call a public API, and here the DeepSeek case is unusually well documented. Its privacy policy states plainly that personal data is collected, processed and stored in China. South Korea's data protection commission found that prompts had been transferred to third party Chinese companies without user consent. Italy's Garante imposed a ban within days of reviewing its practices, investigations followed across more than a dozen European jurisdictions, and the European Data Protection Board stood up a dedicated AI enforcement task force. Several governments restricted the app on official devices, and multiple US states blocked it from state networks.

For a UK or EU business, the practical reading is stark. There is no adequacy decision for China and no documented standard contractual clauses covering deepseek.com, which means routing personal data to the hosted service is not a grey area under GDPR Article 46. It is a transfer without a lawful basis. And this problem is not unique to one country of origin: any public API means your commercially sensitive prompts, your draft contracts, your customer records and your internal policy documents are processed on infrastructure you cannot inspect, under retention terms you did not write.

A private deployment removes the question rather than managing it. Because Aphelion runs the model inside your own environment, prompts and documents never leave your governed perimeter. There is no cross border transfer to justify, no third party retention policy to audit, and no risk that a supplier's future training practices quietly change what happens to your inputs. That is also why data enrichment works differently on a private platform: you can ground the model in your most sensitive material precisely because that material is not going anywhere.

"The open weights are the same either way. What differs is who holds the meter and who holds your data. Run the model yourself and both answers are you."

Stuart Smith, CEO, Aphelion AI

What Ownership Actually Changes

There is an underappreciated irony in this week's news. DeepSeek's models are open weight. The capability that just became significantly more expensive to rent did not change at all for anyone running it on their own hardware. The price rise applies to the convenience of somebody else's servers, not to the intelligence itself.

That is the gap Aphelion is built to occupy. Running a capable model privately converts a variable, externally controlled cost into a fixed one you set, and it changes the incentives that sit underneath everyday decisions. On a metered API, teams ration context, avoid long documents and think twice before letting an agent run a multi step task, because every one of those choices has a price attached. On a flat per user deployment, the correct engineering decision and the cheap decision are the same decision.

Model agnosticism is the other half of the protection. Aphelion is not built around one supplier's weights, so a model that becomes uneconomic, unsupported or unsuitable can be replaced without rebuilding the workflows, prompt library and system integrations layered on top of it. The lock-in that makes public API price rises so effective is an architectural choice, and it is one you can decline to make.

Ten Users, Three Years

Here is the comparison in concrete terms for a ten person team using AI as a working tool rather than an occasional novelty. Aphelion is charged at a flat £25 per user per week, which is £13,000 a year with no build invoice. The public API column assumes moderate agentic use at current published rates, before this week's increase is applied. Figures are indicative, and data enrichment and integrations are scoped and priced to each customer's needs in both models.

The comparison that matters is not the first month. It is what happens in month eighteen, when the tool has become part of how people work and the supplier changes its rate card. One of those columns moves. The other does not.

Being Honest About What Private AI Asks of You

We are not going to pretend the trade is free, because businesses that are told it is free tend to discover otherwise at the worst moment. Private deployment means real hardware, and hardware sets a ceiling. A 1.6 trillion parameter frontier model of the kind DeepSeek now serves at the top of its range is not something a mid sized business runs in a cupboard, and we would not claim otherwise. What runs excellently on sensible private infrastructure is the tier of open weight models that handles the overwhelming majority of genuine business work: drafting, extraction, classification, summarisation, policy lookup, research assistance and structured reasoning over your own data.

The practical question is therefore not "can private AI match the largest frontier model on a benchmark", but "what proportion of your actual work needs it". In most organisations the honest answer is a small fraction. Cost research this year makes the same point from the other direction: organisations that route every workload to frontier models pay many times more per million tokens than those matching model tier to task, and the difference is architectural rather than technical.

Where genuine frontier capability is required, a model agnostic platform lets you route that narrow slice deliberately rather than by default, with full visibility of what it costs. That is a considered decision about a specific workload, not a standing arrangement where every routine request is billed at premium rates on someone else's meter. Applying AI practically to grow a business means being clear eyed about both what it can do and what it should cost, which is the approach the team describes on our About page.

Frequently Asked Questions

Why did DeepSeek raise its V4 API prices?

DeepSeek confirmed on 13 August 2026 that API rates for its V4-Pro and V4-Flash models would rise by between 50 and 1,100 percent depending on the model, the token type and the time of day, with the new rates effective from 17 August. The change also introduces peak and off-peak billing, so the same request costs different amounts depending on when it runs. The wider reason is structural rather than specific to one company: inference costs real money every time a model is called, subsidised launch pricing across the industry is being corrected upward, and providers are moving toward sustainable unit economics. Any business whose AI runs on a rented meter should expect this pattern to repeat.

How does Aphelion AI protect a business from AI price increases?

Aphelion AI runs models privately, inside infrastructure you own or exclusively control, and charges a flat fee of £25 per user per week rather than billing per token. That covers unlimited projects, chats, the prompt library, the prompt builder, policy document to markdown conversion, new features and backups. Because there is no meter, a supplier repricing its API overnight does not change your invoice, and your costs scale with headcount rather than with how heavily your team uses the system. Aphelion is also model agnostic, so if a particular open weight model becomes uneconomic or unsuitable, it can be swapped without rebuilding the workflows sitting on top of it.

Is private AI more secure and compliant than using a public AI API?

In almost every regulated case, yes. When you call a public API, your prompts, documents and outputs leave your environment and land on infrastructure you cannot inspect, in a jurisdiction you did not choose. Hosted DeepSeek is the clearest illustration: its own privacy policy states that personal data is processed and stored in China, Italy's Garante imposed a ban within days of reviewing its practices, investigations opened across more than a dozen European jurisdictions, and multiple governments restricted the app on official devices. A private deployment removes the transfer entirely. Prompts stay inside your governed environment, retention is set by your policy, and a GDPR or ISO audit becomes a review of your own controls rather than a vetting exercise on a third party.

Can a self-hosted AI platform integrate with existing business systems?

It can, and this is usually where the value is created rather than in the raw model. Aphelion connects to the CRMs, ERPs, databases, document stores and line of business systems you already run, so the AI works on your real operational context instead of generic public knowledge. Data enrichment and system integrations are scoped and priced separately from the per user fee, because every estate is different and pretending otherwise produces bad projects. The important structural point is that with a private deployment, feeding the model more context does not increase a per token bill, so richer grounding costs you engineering effort once rather than a permanent tax on every request.

Private AI vs public AI API: which is better value for a business?

A public API wins on speed of experiment and on very low, occasional volume, where a few pounds a month buys frontier capability with no commitment. A private deployment wins the moment AI becomes operational. Public pricing scales with usage and can be changed by the supplier with days of notice, while a flat per user private deployment converts that into a fixed, budgetable line that does not move when your team gets good at using it. Ownership also removes the switching risk that catches most buyers out, because the workflows, prompts and integrations belong to you rather than to a vendor whose pricing page you do not control.

The Meter Is the Message

The lesson of 13 August is not that DeepSeek became expensive. It is that a capability thousands of businesses had already built into their operations was repriced by up to 1,100 percent by a company they have no relationship with beyond an API key, effective in four days. The models did not change. The billing did.

Aphelion exists because that arrangement is a poor foundation for anything a business actually depends on. Run the model privately and the cost is a line you set, the data stays inside your walls, and the capability is an asset on your balance sheet rather than a subscription that can be repriced while you sleep. When the next announcement lands, and there will be a next announcement, the only question worth asking is whether it changes your invoice at all.