DeepSeek Pricing Guide: Chat, API and Model Costs

DeepSeek Chat is currently described as free, while the official developer API is billed by tokens. On 2026-07-30, the API offered V4 Flash and V4 Pro with separate cache-hit input, cache-miss input and output rates. Local open-weight use has no first-party token bill but still has hardware and operating costs.

Verified API prices

Prices are in US dollars per one million tokens.

Source: DeepSeek Models & Pricing↗ — official external destination, checked 2026-07-30. The official page says prices may change. Reverify these numbers before publication, budgeting or implementation.

What each price means

Input tokens are the instructions, conversation history, tool definitions and content sent to the model. A cache miss is input the service must process normally. A cache hit is a previously persisted matching prefix recognised by DeepSeek’s context-caching system. Output tokens include generated material; thinking or reasoning tokens may also contribute to usage.

Do not estimate every request at the cache-hit rate. Prefix matching and persistence determine which tokens qualify. The API response reports prompt cache hit and miss counts, allowing actual usage to be reconciled.

Flash versus Pro

DeepSeek positions V4 Flash as the faster and more economical model and V4 Pro as the stronger model for harder work. Both currently list a one-million-token context length, a maximum output of 384,000 tokens, JSON output, tool calling, thinking and non-thinking modes.

Price is only one selection factor. A lower rate does not save money if the workload needs repeated retries or extensive human repair. Test a representative task set and track success rate, total tokens, latency and review time. The model choice guide provides a practical decision process.

Is DeepSeek Chat included in API pricing?

No. Hosted DeepSeek Chat is a consumer-facing service described as free. The API is a separate platform with keys, balances and token billing. A chat login and an API key may use the same account system, but they are different access methods.

Free chat does not create an unlimited API allowance. Likewise, topping up API balance does not establish an official consumer “Pro” subscription. No such first-party plan was identified when the Pro plan guide was verified.

Balances and promotional credit

DeepSeek’s balance endpoint separates topped-up balance from granted balance. The pricing page says granted balance is deducted first when both are available. However, the existence of that field does not prove that every account gets a promotion.

No universal free-credit amount was documented on the official pricing or balance pages on the research date. If a dashboard shows a granted balance, use its displayed expiry and terms. Do not rely on old referral posts or screenshots.

Context and concurrency

Both current API models list a one-million-token context window and 384,000 maximum output. These are limits, not a recommendation to fill every request. Large prompts cost more and can make relevant details harder to retrieve.

The official rate-limit page listed account-level concurrency of 2,500 for V4 Flash and 500 for V4 Pro. Requests over the limit can receive HTTP 429. These figures are not tokens-per-minute quotas and may be subject to capacity arrangements. Design retries with jitter and respect returned errors.

Local and third-party costs

Open-weight DeepSeek models can be downloaded and run through a local runtime. The weights may not carry a per-token DeepSeek charge, but storage, RAM or VRAM, electricity, backups, staff time and security still create a total cost. Full-size models require very different infrastructure from smaller distilled checkpoints.

Third-party providers set their own rates and limits. Confirm the exact checkpoint, quantisation, context, retention policy and support level. A provider’s “DeepSeek plan” or introductory credit should not be reported as a first-party DeepSeek offer.

Example monthly estimate

Suppose an application uses V4 Flash for 100,000 requests. Each request averages 2,000 uncached input tokens, 3,000 cached input tokens and 800 output tokens.

• Uncached input: 200 million × $0.14 / million = $28
• Cached input: 300 million × $0.0028 / million = $0.84
• Output: 80 million × $0.28 / million = $22.40
• Estimated model total: $51.24

This is an estimate, not a quote. Retries, tool loops, reasoning, larger history and changing rates can increase cost. Use the API cost calculator guide to adapt the formula.

Cost controls

Set output caps appropriate to the task, trim repeated history and order prompts so stable prefixes can be reused. Use Flash for eligible high-volume work and reserve Pro for tasks that demonstrate a quality benefit. Add per-user budgets, alerts and hard limits outside the model.

Log token counts and model IDs without storing unnecessary prompt content. Reconcile invoices or balance changes against application usage. Recheck pricing after model migrations because retired aliases and new releases can change both behaviour and economics.

Conclusion

DeepSeek pricing is simple at the headline level—free hosted chat, paid API and self-funded local compute—but API cost depends on cache status, output volume and model choice. Store current rates centrally, measure real usage and reverify the official pricing page before every publication or budget cycle.