Current DeepSeek API Pricing 2026: V4 Flash & Pro Cost per 1M Tokens
Current DeepSeek API pricing for 2026: V4-Flash-0731 and V4 Pro token costs, cache-hit pricing, the announced peak-hour multiplier, official model IDs, and coding-agent cost examples.
DeepSeek API pricing in 2026 is one of the cheapest options for high-volume AI agents. The current V4 lineup is DeepSeek V4 Flash and DeepSeek V4 Pro, both with 1M context, 384K max output, OpenAI-compatible API support, Anthropic-compatible API support, JSON output, and tool calls.
If you are searching for the current DeepSeek API price, the short answer is: V4 Flash costs $0.0028/M cached input, $0.14/M cache-miss input, and $0.28/M output. V4 Pro costs $0.003625/M cached input, $0.435/M cache-miss input, and $0.87/M output. These are now the permanent base rates.
The important detail is cache pricing. DeepSeek splits input tokens into cache hit and cache miss buckets, and cache-hit input can be 50x cheaper than cache-miss input on V4 Flash.
Official sources: DeepSeek model pricing, DeepSeek updates, DeepSeek context caching, and DeepSeek Anthropic API compatibility.
Quick Answer: DeepSeek V4 API Pricing Per 1M Tokens
If you only need the current official 2026 price table, use this:
| Model | Cached Input / 1M | Cache-Miss Input / 1M | Output / 1M | Context | Best Fit |
|---|---|---|---|---|---|
| deepseek-v4-flash | $0.0028 | $0.14 | $0.28 | 1M | High-volume chat, extraction, and coding-agent subtasks |
| deepseek-v4-pro | $0.003625 | $0.435 | $0.87 | 1M | Harder coding, reasoning-heavy agents, and long-horizon tasks |
DeepSeek V4 also supports OpenAI-compatible and Anthropic-compatible endpoints, JSON output, tool calls, chat prefix completion, and FIM completion in non-thinking mode.
For the broader model page that ranks for DeepSeek V4 pricing, see DeepSeek V4 model pricing. For agent setup, see DeepSeek V4 agent setup.
Current DeepSeek V4 Pricing
All prices below are USD per 1 million tokens.
| Model | Cached Input | Cache-Miss Input | Output | Context | Notes |
|---|---|---|---|---|---|
| deepseek-v4-flash | $0.0028 | $0.14 | $0.28 | 1M | Serves the V4-Flash-0731 public-beta update at the same price |
| deepseek-v4-pro | $0.003625 | $0.435 | $0.87 | 1M | Permanent base price; the 0731 Pro update is expected in early August |
DeepSeek discontinued the deepseek-chat and deepseek-reasoner compatibility aliases on July 24, 2026. For current integrations, use deepseek-v4-flash or deepseek-v4-pro directly.
Is DeepSeek V4-Flash-0731 a New Model?
It is a new checkpoint, but not a new public API identifier. On July 31, DeepSeek upgraded deepseek-v4-flash to the DeepSeek-V4-Flash-0731 public beta under the same architecture, model size, context limits, and token price. The update focuses on stronger agent behavior, Responses API support, and better adaptation to Codex-style workflows.
DeepSeek also announced that both V4 models will eventually use 2x billing during Beijing-time peak windows of 09:00-12:00 and 14:00-18:00. The pricing page says the effective date is still to be determined. This guide and the DevTk.AI calculator therefore continue to use the current base price until DeepSeek publishes that date.
Why Real Bills Can Be Much Lower Than the Table Price
DeepSeek’s cost formula is:
cost =
prompt_cache_hit_tokens * cached_input_price / 1,000,000
+ prompt_cache_miss_tokens * cache_miss_input_price / 1,000,000
+ completion_tokens * output_price / 1,000,000The cache is enabled by default. You do not need to change application code to use it. DeepSeek exposes two usage fields so you can measure the effect:
usage.prompt_cache_hit_tokens
usage.prompt_cache_miss_tokensCache hits are best-effort, not guaranteed. They work best when requests reuse a stable prefix, such as the same system prompt, coding rules, repository summary, product spec, or previous conversation prefix.
How 10M Tokens Can Cost Around RMB 2
The “10 million tokens for a little over RMB 2” story is plausible when most input is cached and output is small. Using DeepSeek’s official domestic V4 Flash prices:
8.5M cached input * ¥0.02/M = ¥0.17
1.0M cache-miss input * ¥1/M = ¥1.00
0.5M output * ¥2/M = ¥1.00
Total = ¥2.17That is not the cost of 10M random uncached tokens. It is the cost of a cache-heavy workflow where repeated context dominates, which is exactly what long-running coding agents often do.
Cost Examples
| Workload | No Cache, V4 Flash | 85% Input Cache Hit, V4 Flash |
|---|---|---|
| 10M input + 1M output | $1.68 | $0.49 |
| 50M input + 5M output | $8.40 | $2.45 |
| 500M input + 50M output | $84.00 | $24.50 |
For repeated agent work, cache hit rate matters more than the headline input price. Monitor the prompt_cache_hit_tokens ratio before deciding whether V4 Flash, V4 Pro, or another provider is cheaper.
DeepSeek V4 Pro Permanent Price
V4 Pro’s former 75%-off promotion ended on May 31, 2026 at 15:59 UTC. DeepSeek retained the discounted amount as the permanent price, equal to one quarter of the original reference rate.
| Price Type | Current Permanent Price | Original Reference Price |
|---|---|---|
| Cached input | $0.003625/M | $0.0145/M |
| Cache-miss input | $0.435/M | $1.74/M |
| Output | $0.87/M | $3.48/M |
Use V4 Pro when you need stronger agentic reasoning, complex coding, or long-horizon task coherence. Use V4 Flash for high-volume production flows where cost and latency dominate.
DeepSeek API Free Tier, Credits, and Balance
DeepSeek’s public API docs do not publish a permanent free-tier table for V4. The billing docs say usage is deducted from your topped-up balance or granted balance, with granted balance used first when both are available.
For planning, treat DeepSeek V4 as pay-as-you-go unless your account dashboard shows active granted credits. If you are checking whether you have free credits, use the platform balance page or the /user/balance API response fields:
total_balance
granted_balance
topped_up_balanceThe practical takeaway: DeepSeek may show granted balance on some accounts or promotions, but do not assume ongoing free production capacity unless your own account balance confirms it.
OpenAI-Compatible API Example
from openai import OpenAI
client = OpenAI(
api_key="your-deepseek-api-key",
base_url="https://api.deepseek.com"
)
response = client.chat.completions.create(
model="deepseek-v4-flash",
messages=[
{"role": "system", "content": "You are a careful coding assistant."},
{"role": "user", "content": "Explain this repository structure."}
],
)
print(response.choices[0].message.content)
print(response.usage.prompt_cache_hit_tokens)
print(response.usage.prompt_cache_miss_tokens)For coding agents that expect Anthropic format, use:
export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=your-deepseek-api-keyWhen DeepSeek Is the Best Choice
| Use Case | Recommendation |
|---|---|
| High-volume chat, extraction, summaries | Start with V4 Flash |
| Coding agents with repeated repo context | Start with V4 Flash, monitor cache hit rate |
| Harder autonomous agent tasks | Use V4 Pro when quality matters more than the lowest possible token cost |
| Vision or audio input | Use another provider; DeepSeek V4 is text-first |
| Enterprise compliance-sensitive data | Evaluate legal and data-residency requirements first |
Related guides:
- How to Configure DeepSeek V4 in Claude Code — terminal agent setup
- DeepSeek V4 for OpenCode, Codex, Cline, Kilo, and Roo — multi-agent setup notes
- AI API Pricing Comparison 2026 — compare DeepSeek with OpenAI, Claude, Gemini, MiMo, and others
- AI Model Pricing Calculator — estimate your monthly bill