API Pricing Compared: DeepSeek vs GPT, Claude and Gemini by Real Task Cost

API pricing is usually advertised per million tokens, a unit that is convenient for providers and slightly abstract for everyone else. A million tokens is roughly 700,000–750,000 English words, although the exact number depends on the language and formatting. For a normal article draft, support reply or code review, the bill is usually measured in fractions of a cent to a few cents.

The important distinction is between input tokens—your prompt and attached context—and output tokens, which the model generates. Output is normally more expensive. A long system prompt, conversation history or retrieved document can therefore cost more than the visible question suggests.

Reference API prices

The table uses representative official list prices for commonly used API models. Prices are shown in USD per one million tokens, with an indicative EUR conversion at €0.92 per US dollar. The EUR figures are rounded and exclude VAT, payment fees and any provider-specific taxes. Currency markets move; your card statement may disagree by a few percent.

For the primary prices, check the providers’ live documentation before committing budget: DeepSeek pricing, OpenAI API pricing, Anthropic API pricing and Google Gemini API pricing. Providers frequently add aliases, retire models or change regional billing rules.

What does one typical task cost?

To make the numbers usable, I calculated a task containing 10,000 input tokens and 2,000 output tokens. That is a reasonable approximation for summarising a long document, producing a structured article draft, reviewing a medium-sized code change or answering a support ticket with retrieved context.

The calculation is simple:

Cost = (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price)

These figures are deliberately mechanical rather than a claim that every model produces the same quality. A Sonnet response that solves a problem in one attempt may be cheaper than three failed attempts with a cheaper model. The spreadsheet cannot measure developer frustration, though it should probably have a column for it.

DeepSeek is cheap—but price is not the entire decision

DeepSeek-V3 is among the lowest-cost options in this comparison, while DeepSeek-R1 remains inexpensive for a reasoning-oriented model. This makes DeepSeek attractive for high-volume classification, first-pass drafting, extraction and applications where a small error rate can be handled by validation or a second model.

There are two operational questions for European companies. First, confirm the exact API endpoint and its current data-processing terms. DeepSeek’s public service has been associated with data storage and processing in China, which can create additional GDPR transfer and vendor-assessment work. Second, decide whether your organisation is permitted to send personal, confidential or regulated data to that service. “The API is reachable from the EU” is not the same as “your compliance department has approved the processing arrangement”.

DeepSeek also offers open-weight models that can be self-hosted. That changes the calculation: there is no per-token provider invoice, but GPU hardware, electricity, maintenance, monitoring and engineering time become your costs. On a modest local GPU, a large reasoning model may be technically interesting and commercially impractical. Our AI Arena benchmark work treats local inference as an infrastructure comparison, not as magically free computing.

GPT, Claude or Gemini: when the higher price can make sense

GPT-4.1 mini and GPT-4.1

GPT-4.1 mini is a sensible default for applications that need a mature ecosystem, tool calling and predictable latency without paying flagship rates. GPT-4.1 costs more, but the price can be justified for complex instruction following, code generation and tasks where retries are expensive.

OpenAI provides API availability in Europe, and eligible customers may have access to European data-residency controls. Those controls depend on the product, project and contract, so verify them rather than assuming that an EU billing address equals EU processing.

Claude

Claude Sonnet is expensive on a pure token basis, particularly when the response is long. It is often selected for writing, code review, analysis and nuanced instruction following. Haiku is the more economical choice for routing, extraction and short responses.

Anthropic’s commercial terms and regional options should be checked for the exact API product you plan to use. For GDPR-sensitive workloads, the practical checklist remains the same: sign the relevant data-processing agreement, establish retention rules, determine transfer mechanisms and avoid sending personal data unless the use case requires it.

Gemini

Gemini 2.5 Flash is one of the strongest price candidates for large-volume multimodal or long-context work. Gemini 2.5 Pro costs considerably more, but its pricing can still compare favourably with premium models for large documents and difficult coding tasks.

Google offers Gemini through its developer API and through Vertex AI. Vertex AI is usually the more relevant route for organisations that need cloud-region selection, enterprise controls and integration with existing Google Cloud governance. Availability and specific residency guarantees vary by product and region, including within the EU.

Three costs that token tables hide

1. Cached input and batch processing

Repeated system prompts, product catalogues and long reference documents may qualify for cached-input pricing. Batch APIs can also reduce costs for non-urgent jobs. Before comparing providers, identify how much of your input is repeated and whether the task needs an immediate response.

2. Reasoning and tool tokens

Reasoning models may consume hidden or billed reasoning tokens. Tool calls can add another round of input and output, while web search, image processing, audio and file handling may have separate charges. A “2,000-token answer” is not necessarily a 2,000-token API transaction.

3. Retries and quality control

Budget for validation. In production, I would rather pay for a cheap model plus schema validation and a targeted retry than trust an unverified paragraph because its first response cost €0.005. For important outputs, a two-stage pipeline—small model for triage, stronger model for difficult cases—usually beats sending everything to the premium tier.

A practical selection method for European teams

1. Log real traffic. Record input tokens, output tokens, model, latency, retries and tool usage for at least several hundred representative requests.
2. Build a quality test set. Include your actual languages, documents and failure cases. English-only public benchmarks are not enough for Czech, German, French or mixed EU content.
3. Calculate total task cost. Add retries, caching, batch discounts, storage, observability and VAT where applicable.
4. Check data handling. Confirm training opt-out, retention, subprocessors, transfer mechanisms and available EU regions with the provider’s current legal documentation.
5. Route by difficulty. Use DeepSeek-V3, GPT-4.1 mini or Gemini Flash for routine work; reserve R1, GPT-4.1, Claude Sonnet or Gemini Pro for cases that actually need them.

For GDPR, the provider is normally a processor only when the contractual arrangement and processing actually support that role. Your company remains responsible for lawful collection, purpose limitation, minimisation, access control and deletion. The EU AI Act adds obligations according to the system’s use and risk category; it does not turn a low token price into a compliance exemption.

Bottom line

On this representative task, DeepSeek-V3 is the cheapest, followed closely by GPT-4.1 mini and Gemini 2.5 Flash. DeepSeek-R1 remains inexpensive for reasoning. Claude Sonnet 4, GPT-4.1 and Gemini 2.5 Pro cost several times more, but can be economically sensible if they reduce retries or human review.

My recommendation is to treat published token prices as the starting line, not the buying decision. Measure your own prompts, test European-language quality, and keep sensitive data in a region and contractual setup your organisation can defend. A €0.005 request is cheap; a badly governed production system is not.