DeepSeek API Pricing (Oct 2026): V4.1 Flash vs V4 Pro

DeepSeek's official API charges $0.15 per 1M input tokens and $0.60 per 1M output tokens for V4.1 Flash off-peak, and exactly double ($0.30 / $1.20) during peak hours. V4 Pro is $0.66 / $1.98 off-peak and $1.32 / $3.96 at peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. On Orux AI, deepseek-v4.1-flash costs 22.92 / 91.67 Credits ($0.115 / $0.458) off-peak, but Orux applies the peak rate every day of the week, which makes weekend daytime calls more expensive than going direct. Details, tables and Python below.

DeepSeek API prices in October 2026

Official prices per 1M tokens, from DeepSeek's pricing page, as of Oct 29, 2026:

Swipe horizontally to view the full table.

DeepSeek lists a 1M-token context and up to 384K output tokens for both models, with JSON output and tool calls supported.

The same models on Orux AI, per 1M tokens, as of Oct 29, 2026. Orux AI bills in Credits at a fixed 1 Credit = $0.005; dollar values are only for comparison.

Swipe horizontally to view the full table.

Two practical notes on that table. First, use deepseek-v4.1-flash, not deepseek-v4.1-flash-1017. The 1017 listing has a flat price equal to DeepSeek's peak rate around the clock, so off-peak it costs twice the official price and about 2.6 times the Orux off-peak rate for deepseek-v4.1-flash. Second, every chat request on Orux AI has a minimum charge of 0.001 Credits, which only matters for very short calls.

Peak and off-peak hours in UTC

DeepSeek defines peak as 09:00 to 12:00 and 14:00 to 18:00 Beijing time (UTC+8). In UTC that is:

Swipe horizontally to view the full table.

This is where Orux AI is more expensive. DeepSeek applies peak pricing only on weekdays, excluding Chinese public holidays. Orux AI applies its peak band every day. During those weekend and holiday hours, deepseek-v4.1-flash on Orux costs $0.229 / $0.917 per 1M tokens while DeepSeek charges its off-peak $0.15 / $0.60, so Orux is about 53% more. For V4 Pro the gap in those hours is larger: $1.125 / $3.375 on Orux versus $0.66 / $1.98 direct.

Averaged over a full week of evenly spread traffic, Orux AI still comes out lower, because its off-peak rate is well under DeepSeek's. For the workload in the next section, a week of flat traffic averages about 118.4 Credits ($0.592) per 1,000 Flash requests on Orux versus about $0.725 direct, ignoring holidays. But if your jobs cluster on weekend mornings in Europe or weekend afternoons in Asia, run the numbers for your own schedule.

V4.1 Flash vs V4 Pro: which one

Cost per 1,000 requests, each with 2,000 input and 500 output tokens and no cache hits, as of Oct 29, 2026:

Swipe horizontally to view the full table.

V4 Pro costs about 4.3 times as much as Flash for the same tokens. Our rule of thumb is to start every workload on Flash, build a small set of graded examples, and move only the prompts that fail those examples to Pro. Summaries, classification, extraction and short copy rarely need Pro. Multi-step reasoning over long documents and harder code are where the extra cost has a chance to pay back, but that is something to measure on your own examples, not assume. For cheaper models from other vendors on the same workload, see cheapest LLM API.

Call DeepSeek from Python

Orux AI exposes an OpenAI-compatible endpoint, so the official OpenAI SDK works with only base_url changed. Create a key at /keys; it is shown once.

# pip install openai
import json, os
from datetime import datetime, timezone
from openai import OpenAI

client = OpenAI(base_url="https://orux.top/api/v1", api_key=os.environ["ORUX_API_KEY"])

resp = client.chat.completions.create(
    model="deepseek-v4.1-flash",
    messages=[{
        "role": "user",
        "content": "Return JSON with keys name (string), opens (YYYY-MM-DD) and hours (string) "
                   "for: Corner Cafe opens on 2026-11-08, open 08:00-18:00.",
    }],
    response_format={"type": "json_object"},
    max_tokens=200,
)
data = json.loads(resp.choices[0].message.content)
print(data)

# Orux AI rates for deepseek-v4.1-flash, Credits per 1M tokens, checked 2026-10-29
RATES = {"offpeak": (22.9167, 91.6668), "peak": (45.8334, 183.3336)}
hour = datetime.now(timezone.utc).hour
band = "peak" if (1 <= hour < 4 or 6 <= hour < 10) else "offpeak"  # every day on Orux AI
inp, out = RATES[band]
u = resp.usage
print(band, f"~ {(u.prompt_tokens * inp + u.completion_tokens * out) / 1e6:.6f} Credits")

With response_format={"type": "json_object"}, describe the JSON structure you want in the prompt itself, including the keys and types; the format flag asks for valid JSON, not for your schema. tools is also supported for function calling. Error codes follow the usual pattern: 401 is a bad key, 402 means the balance is empty and needs a top-up, and 429 means back off exponentially with jitter.

Cut costs: schedule batch jobs off-peak

If a job can wait, the cheapest move is to run it outside the peak band. On Orux AI the off-peak rate is half the peak rate, every day. This helper tells you whether a moment is peak on Orux AI or on DeepSeek direct, and when the next Orux off-peak minute starts.

from datetime import datetime, time, timedelta, timezone

# Orux AI peak bands for DeepSeek, every day (UTC). Official peak applies Mon-Fri only.
PEAK_UTC = [(time(1, 0), time(4, 0)), (time(6, 0), time(10, 0))]


def orux_is_peak(now=None):
    t = (now or datetime.now(timezone.utc)).astimezone(timezone.utc).time()
    return any(start <= t < end for start, end in PEAK_UTC)


def official_is_peak(now=None):
    now = (now or datetime.now(timezone.utc)).astimezone(timezone.utc)
    return now.weekday() < 5 and orux_is_peak(now)  # ignores Chinese public holidays


def next_offpeak(now=None):
    """Return the first minute at or after `now` that Orux AI bills at the off-peak rate."""
    t = (now or datetime.now(timezone.utc)).astimezone(timezone.utc).replace(second=0, microsecond=0)
    while orux_is_peak(t):
        t += timedelta(minutes=1)
    return t


if __name__ == "__main__":
    now = datetime.now(timezone.utc)
    print("Orux peak now:", orux_is_peak(now), "| official peak now:", official_is_peak(now))
    print("Next Orux off-peak minute:", next_offpeak(now).isoformat())

For example, at 07:30 UTC on Saturday, Oct 31, 2026, orux_is_peak returns True, official_is_peak returns False, and next_offpeak returns 10:00 UTC. Put a check like this in front of a nightly or backlog queue: if it is peak, sleep until next_offpeak(). Interactive traffic should not wait, but a 50,000-document summarization run usually can, and it halves the bill.

Case study: stop the model inventing offers

Cheap tokens do not help if the output is unusable. We ran a small Chinese copywriting test for a fictional café. The calls ran on the deepseek-v4.1-flash-1017 listing with max_tokens=512, temperature=0.7 and reasoning_effort="none". These are two inspected outputs, not an accuracy benchmark.

The first request asked for warm, short opening copy of no more than 100 characters, with no facts supplied. The response (recorded Oct 27, 2026) was fluent and 75 characters long, and it invented two things: "opening today" and a free first cup ("第一杯，算我的", roughly "the first cup is on me"). It also named a latte the brief never mentioned. None of that is a style problem; each is an unapproved claim a café owner would have to honor.

The rerun (Oct 29, 2026) used a brief with two halves: facts that must appear, and claims that must not. The exact prompt:

请为下面的虚构社区咖啡店写一段可以直接放在开业海报上的中文文案。只输出一段文案，不加标题和解释。
已确认事实：店名「转角咖啡」；开业日期 2026年11月8日；地址「青禾路18号」；营业时间 08:00–18:00；供应咖啡和茶。
要求：温暖、平实；必须包含店名、完整日期、地址和营业时间；不超过100个Unicode字符，标点、数字和空格计入，不计换行。
没有获批的优惠活动，不得添加免费、赠送、折扣、首杯优惠；不得写「今日开业」，不得添加其他产品、价格、承诺或未经提供的事实。

In English: the brief gives the name (Corner Café), the opening date (November 8, 2026), the address (18 Qinghe Road), hours (08:00 to 18:00) and the menu (coffee and tea); it requires all four details within 100 Unicode characters; and it forbids free items, discounts, giveaways, "opening today" and any other unsupplied facts.

The actual response:

2026年11月8日，转角咖啡在青禾路18号开业，08:00–18:00营业，供应咖啡和茶。

Translated for reading: "Corner Café opens at 18 Qinghe Road on November 8, 2026. Open 08:00 to 18:00, serving coffee and tea." It is 46 characters, contains all four required details exactly as supplied, adds no offer, and is rather plain. Warmth is an editing pass that comes after the fact check, not before it.

The rerun used 154 input and 32 output tokens and finished with stop. What it cost:

Swipe horizontally to view the full table.

The lesson carries over to any model: a brief with a "must include" list and a "must not add" list is cheaper than any amount of post-editing. We have not rerun this test on the deepseek-v4.1-flash listing, so the two middle rows are price calculations, not observed runs.

FAQ

Q.How much does the DeepSeek API cost?

DeepSeek V4.1 Flash costs $0.15 per 1M input tokens and $0.60 per 1M output tokens off-peak, and $0.30 / $1.20 at peak, at DeepSeek's official prices as of Oct 29, 2026. V4 Pro is $0.66 / $1.98 off-peak and $1.32 / $3.96 at peak. On Orux AI, V4.1 Flash is 22.92 / 91.67 Credits ($0.115 / $0.458) off-peak.

Q.When are DeepSeek off-peak hours?

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC (09:00 to 12:00 and 14:00 to 18:00 Beijing time). At DeepSeek directly, peak applies Monday to Friday only, excluding Chinese public holidays. On Orux AI, the same peak band applies every day, including weekends. All other hours are off-peak.

Q.Is DeepSeek V4 Pro worth it over Flash?

Only for prompts that Flash demonstrably fails. V4 Pro costs about 4.3 times as much per request: 393.75 versus 91.67 Credits per 1,000 typical requests off-peak on Orux AI. Start on V4.1 Flash, keep a set of graded test prompts, and route only the failures to Pro.

Q.Does cache hit pricing apply automatically?

At DeepSeek, context caching is enabled by default: repeated prompt prefixes are served from cache and billed at the cache-hit rate without code changes. On Orux AI, cache-hit tokens reported in the response usage are billed at the cache-read rate, 0.458 Credits per 1M off-peak for V4.1 Flash, about 50 times cheaper than regular input.

Q.Can I use the OpenAI SDK with DeepSeek?

Yes. Set base_url="https://orux.top/api/v1", use your Orux AI key, and pass model="deepseek-v4.1-flash" or model="deepseek-v4-pro-0813". Standard Chat Completions parameters work, including response_format={"type": "json_object"} and tools.

Sources and next step

Try the saved examples on the DeepSeek V4.1 Flash model page, and for a wider Claude cost comparison see Claude API pricing. Prices checked on 2026-10-29. Sources: DeepSeek pricing, DeepSeek context caching, Orux AI pricing.