How to Run DeepSeek V4 Flash via API in 2026 (Priced)

DeepSeek V4 Flash is the cheapest frontier-class open checkpoint on the market: DeepSeek listed it at $0.14 per million input tokens and $0.28 per million output at its official API on 2026-08-20, and the same open weights are served by nearly every major inference host — often for less. This run-guide covers what the model actually is, who serves it at what price, how cache-hit billing changes your real bill, and the exact steps to switch.

DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.

Key facts

• Official list price (2026-08-20): DeepSeek's API prices V4 Flash at $0.14 input / $0.28 output per million tokens, with cache-hit input billed at a discount — the single cheapest frontier-class rate DeAI has verified to date, per DeepSeek's pricing page.
• Open weights: the checkpoint ships for self-hosting, so the API is one option among many — not a monopoly.
• Third-party spread: the same weights appear on Fireworks, Together, DeepInfra, Groq, Cerebras, SambaNova, Chutes, and OpenRouter at prices that vary by host and week; on one September snapshot, an aggregator listed the model as low as $0.037 input before a catalog revision (DeAI Price Index).
• Cache economics are the real differentiator: DeepSeek's official API discounts cache-hit input tokens, so a workload with long, repeated system prompts can pay a fraction of the sticker input rate.
• License: DeepSeek's open-weight license permits commercial use and self-hosting; read the terms before bundling it into a product.

What DeepSeek V4 Flash actually is

DeepSeek V4 Flash is the fast, cheap tier of DeepSeek's V4 generation — a mixture-of-experts open-weights model from Chinese AI lab DeepSeek, positioned as the high-volume workhorse beneath the V4 Pro flagship. The design goal is throughput: fewer active parameters per token, a shorter reasoning budget than Pro, and a per-token price tuned for workloads where you run millions of tokens per hour — batch summarization, classification, extraction, agent loops.

The Flash/Pro split mirrors what every major lab now ships: one flagship checkpoint that chases benchmark ceilings and one workhorse checkpoint that chases $/M tokens. For DeepSeek, the published anchor is price: $0.14/$0.28 per million tokens on the official API as of 2026-08-20 (DeepSeek pricing). V4 Pro, the flagship, commands a multiple of that rate; on the same date DeAI recorded it as the frontier pick for code-heavy workloads (DeAI cites its V4 Pro run-guide for that comparison).

DeepSeek publishes the checkpoint as open weights, which is why this guide is not just "sign up for the official API": the same weights are served by third-party hosts at prices that sometimes undercut the official rate and sometimes don't. Where a host lands in that spread is the whole decision.

Who serves V4 Flash, and at what price

Prices below are list rates observed on provider pricing pages as of 2026-09-24 and move weekly; confirm on the day you buy. Where a provider doesn't publish a per-model list, we say so rather than invent a number.

Two rows in that table deserve their own note.

The official API is the price anchor, not necessarily the price floor. DeepSeek's $0.14/$0.28 is the number other hosts get compared against, but third-party GPU serverless hosts have been observed listing the same checkpoint below it — OpenRouter's catalog, for instance, showed rates as low as $0.037 input earlier in September before a revision that DeAI's Price Index delta flagged as possible catalog data noise. The lesson is not that one host is permanently cheapest; it's that this checkpoint's price spreads across hosts are wide enough to justify a dated check before you commit spend.

The TEE row is a different product. Chutes serves the same weights inside hardware-attested trusted execution environments on Bittensor, where the claim is that GPU operators cannot read your prompts — an attestation/policy claim DeAI has not independently verified. It is not a price play and not a speed play; it exists for workloads where verifiable execution, not cost, is the requirement. For the broader category, see the decentralized AI inference networks that actually work.

Cache-hit pricing: the number that actually matters

DeepSeek's API prices cache-hit input tokens at a discount to the base rate — the exact multiplier is published on the pricing page and has changed before, so treat the current card as authoritative. The economics are worth understanding because they change which host is cheapest for your traffic, not in the abstract:

1. What gets cached: the KV cache of your prompt prefix. A long system prompt that is identical across thousands of requests is the canonical cache hit; a unique user query appended to it is not.
2. What gets billed at full rate: every token of the prompt that misses the cache, plus every output token.
3. The workload test: if your prompts share a long, stable prefix — the same system prompt, the same retrieved context — a cache-discounted host can effectively halve your input bill. If every request is unique end to end, the cache discount buys you almost nothing and the base $/M rate is the whole story.

Before you budget against any row in the table above, compute your own input/output split and cache-hit ratio from logs, then re-run the comparison. The sticker rate is a starting point, not the answer.

Switching to V4 Flash: the afternoon version

The official DeepSeek API and nearly every third-party host expose an OpenAI-compatible chat-completions endpoint, so the migration is a base-URL and API-key change:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com/v1",  # swap per provider
    api_key=os.environ["DEEPSEEK_API_KEY"],
)

resp = client.chat.completions.create(
    model="deepseek-chat",  # exact slug varies by provider
    messages=[{"role": "user", "content": "ping"}],
)

Keep the base URL and model slug in config, not code, and a host change becomes a deploy-time decision. Three cautions from the field:

1. Model slugs are not portable. The official API calls it deepseek-chat; Fireworks uses deepseek-ai/DeepSeek-V4-Flash; Chutes appends a -TEE suffix; OpenRouter namespaces it under deepseek/. Map exact names before cutover.
2. Sampling defaults differ per host. Temperature handling, max-token clamping, and serving stack (quantization, batching) all vary; the same weights can score differently on your evals across hosts.
3. A cheaper host that fails your evals is not cheaper. Run an A/B on your own traffic — a canary at 5% for a couple of hours — before cutting over a production workload. The full mechanics are in how to migrate off the OpenAI API and what an OpenAI-compatible API actually is.

Self-hosting: when the open weights beat the API

The open weights are a legitimate escape hatch — and sometimes the cheapest option at scale. Run the checkpoint yourself with vLLM or SGLang behind an OpenAI-compatible server, and your marginal token cost becomes electricity and amortized hardware. The tradeoffs are real: the full model needs multi-GPU VRAM you will not fit on a single consumer card, serving it well is an ops discipline (quantization choice, batching, uptime), and you inherit the security burden — exposed self-hosted inference endpoints are a recurring incident class DeAI tracks, most recently in a 150,000-endpoint internet scan.

The honest framing: self-hosting wins when you have steady, predictable volume, spare GPU capacity, and in-house ops; hosted APIs win when traffic is spiky, when you want zero maintenance, or when you need a checkpoint you can't serve locally. DeAI ran this math in detail in self-hosting vs API cost, and the DeepSeek official-vs-third-party comparison covers the jurisdiction angle — DeepSeek is a Chinese lab, and data-residency requirements push some buyers toward non-China hosts of the same open weights.

Privacy: read the retention terms, not the marketing

DeepSeek's published terms have historically allowed use of inputs to improve its services, with an opt-out available in account settings — a policy DeAI has not independently verified, and one that has changed before. Third-party hosts each publish their own retention terms, and a host can be a better privacy fit than the official API even when serving the same weights. For prompts that must stay confidential by construction, the TEE-attested option in the table is the structural answer, with the standing caveat that "operators can't see prompts" is an attestation claim, not an independent verification. DeAI's Provider Trust Tracker tracks retention, logging, and jurisdiction per provider.

FAQ

How much does the DeepSeek V4 Flash API cost?

DeepSeek's official API listed V4 Flash at $0.14 per million input tokens and $0.28 per million output tokens as of 2026-08-20. Output tokens cost 2× the input rate on the official card, and cache-hit input tokens are billed at a discount — the exact multiplier is on the pricing page and has changed before, so check it on the day you budget.

What is the difference between DeepSeek V4 Flash and V4 Pro?

V4 Flash is the fast, cheap workhorse tier — lower per-token prices, a shorter reasoning budget, and the checkpoint most hosts replicate for bulk traffic. V4 Pro is the flagship: higher benchmark scores and a longer effective reasoning budget at a higher price. Use Flash for high-volume workloads; use Pro where a few points of benchmark accuracy change the outcome.

Can I run DeepSeek V4 Flash on my own GPU?

Yes. The checkpoint ships as open weights, so you can self-host with vLLM or SGLang. The practical floor is multiple high-VRAM GPUs — a single 24 GB card will not hold the full model. If the ops burden isn't worth it, every host in this guide serves the same weights over an API.

Does the DeepSeek API train on my prompts?

DeepSeek's published terms have historically allowed use of inputs to improve services, with an opt-out path in account settings — a policy DeAI has not independently verified. Third-party hosts publish their own retention terms, and TEE-attested networks such as Chutes claim operators cannot see prompts (an attestation claim, not a verification). Check the current policy before sending sensitive traffic.

Can I switch to DeepSeek V4 Flash without rewriting code?

Usually yes. The official API and nearly every third-party host expose an OpenAI-compatible chat-completions endpoint, so switching is typically a base-URL and API-key change. Model slugs are not portable across providers — deepseek-chat, deepseek-ai/DeepSeek-V4-Flash, and namespaced variants all refer to related checkpoints — so map exact names before you cut over.

Current state (September 2026)

V4 Flash has become the default budget workhorse of the open-weights era: it is the checkpoint most often cited in price comparisons, the one most hosts replicate, and the row that shows the widest cross-host price spread in DeAI's weekly snapshots. That spread is the actionable part — the same open weights can cost 3–4× more on one host than another on a given week, and the cache-hit terms can swing the real bill again. Price it fresh on the day you buy, benchmark against your own traffic, and treat every list rate in this guide as a dated observation rather than a standing offer. For the full field, see DeAI's Inference Price Index and The Cheapest LLM APIs in 2026.