DeepSeek Pricing 2026: V4-Flash & V4-Pro API Costs

DeepSeek API pricing at a glance

As of July 2026 the API exposes two models. The legacy names deepseek-chat and deepseek-reasoner were retired on 24 July 2026 at 15:59 UTC. Until that moment they mapped to V4-Flash (non-thinking and thinking modes respectively); since then they are fully retired and inaccessible — calls are no longer routed anywhere. Read what changed on your bill in deepseek-reasoner retired: migrate to V4-Flash, not Pro.

Table 1 — Current DeepSeek API rates, per 1M tokens, USD. Verified 2026-07-25.

Need to wire this up in code? See the DeepSeek API setup guide.

How context caching cuts your bill

DeepSeek automatically caches the prefix of your prompts on disk. When a new request repeats a prefix it has recently seen — a system prompt, a long document, a few-shot example set — those repeated input tokens are billed at the much lower cache-hit rate instead of the cache-miss rate. This is automatic: no SDK changes, no cache-control headers, no opt-in flag. DeepSeek shipped disk-based context caching on 2 August 2024, and it has applied to every model generation since — including V4.

The practical impact is large for any workload with a stable prefix. If 90% of your input tokens are cache hits on V4-Flash, your effective input price drops from $0.14 to roughly $0.017 per 1M — an ~88% saving on input, with no code changes. To maximise it: keep the reusable part of your prompt (system instructions, context documents) at the front, and the variable part at the end.

Peak-hour pricing: announced, not yet active

DeepSeek's off-peak discount — 16:30–00:30 UTC daily, 50% off DeepSeek-V3 and 75% off DeepSeek-R1 — applied to the V3/R1 generation, not to V4. It ended with those models: the deepseek-chat and deepseek-reasoner aliases were retired on 24 July 2026 at 15:59 UTC.

What DeepSeek announced on 30 June 2026 is the opposite. For the official V4 release, rates double during two daily windows:

Outside those windows the rates in Table 1 apply unchanged — a surcharge on peak hours, not a discount off-peak.

Two caveats we'd rather state than hide: the 2× figure and the windows come from DeepSeek's 30 June announcement as carried by the press, not from a live rate card. And the timezone is our reading — the hours are almost certainly Beijing time (UTC+8), but DeepSeek's wording didn't specify.

If it ships as announced, a workload spread evenly across the day costs roughly 29% more; one that avoids both windows costs nothing extra.

Worked examples

How DeepSeek billing works

DeepSeek bills on tokens, not on requests or seats. Every API call is charged for its input tokens (your prompt, system message and any supplied context) plus its output tokens (the model's completion); in thinking mode, reasoning tokens are billed as output. A token is roughly ¾ of an English word — 1M tokens is about 750,000 words. There is no monthly minimum and no per-seat fee: you pay for exactly what you send and receive, deducted post-paid from your prepaid balance.

Rate limits: concurrency only

DeepSeek does not publish requests-per-minute (RPM) or tokens-per-minute (TPM) limits — neither figure appears anywhere in its documentation. The only documented ceiling is concurrency: how many requests you may have in flight at once.

There are no account tiers. Limits do not rise automatically with paid usage or account age — a higher concurrency ceiling requires an explicit capacity expansion request to DeepSeek. If you hit a 429 Too Many Requests, reduce the number of in-flight requests, retry with exponential back-off, and queue rather than fan out. If you need a permanently higher ceiling, contact DeepSeek to request expanded capacity.

V4-Flash vs V4-Pro: which should you use?

Default to V4-Flash. It handles the vast majority of production workloads — chat, RAG, extraction, classification, translation, tool-calling — at a fraction of Pro's cost and with the same 1M-token context window. On typical mixed workloads, Flash is roughly 3× cheaper end-to-end than Pro.

Escalate to V4-Pro only for calls that measurably fail on Flash: multi-step reasoning, competitive-math or code-golf problems, long-horizon agent planning, and heavy self-debugging coding sessions. A common cost-effective pattern is hybrid routing — send every call to Flash first, and re-run on Pro only when a confidence check, evaluation model, or user thumb-down flags the answer as insufficient.

For a deeper capability breakdown, see the DeepSeek V4 model guide.

DeepSeek pricing at scale

The gap widens dramatically at production volumes. A team running 100M tokens/day on V4-Flash with an 80% cache-hit ratio pays roughly $340/month in input + output. The same token volume on a US frontier API runs into the tens of thousands of dollars per month — run your own numbers against the rate table below rather than trusting a headline multiple.

The three levers that matter most for cost control at scale: (1) route to V4-Flash by default, (2) engineer prompts so the reusable prefix comes first for maximum cache-hit rate, and (3) schedule non-latency-critical batch work outside the announced peak-hour window to sidestep the surcharge once it is active. Together those three moves typically cut a naive DeepSeek bill by 60–80%.

DeepSeek pricing vs ChatGPT and Claude (2026)

Input / output rates per 1M tokens, verified 2026-07-25. Competitor rates change frequently — verify each provider's current pricing before committing.

For full feature breakdowns see our DeepSeek vs ChatGPT and DeepSeek vs Claude comparisons.

How we keep this page accurate

Pricing pages are only useful when the numbers are current. This page is maintained by the deepseek.ai editorial desk and is re-verified on a fixed cadence. All rates on this page come from a single rate-config file in our codebase — the table, the calculator and the "last verified" stamp cannot drift out of sync because they all read the same source.

Our verification workflow: on every scheduled review we cross-check three things against DeepSeek's official API pricing page — the standing per-token rates for each model, the cache-hit vs cache-miss columns, and the status of the announced peak-hour surcharge. Any changes go into the rate-config, the "Last verified" date updates, and a note lands in the blog changelog.

deepseek.ai is not affiliated with DeepSeek. If you spot a stale number, email us via the contact page — corrections typically ship the same day.

How to start using DeepSeek

Frequently asked questions