DeepSeek raised API prices on August 16, 2026. Before you budget your next project, here's exactly what V4 Flash costs now — and where solo devs are finding better rates.
DeepSeek V4 Flash costs $0.22 per million input tokens off-peak as of August 16, 2026, up from the previous $0.14 flat rate. For comparable open-weight quality at $0.59/M for Llama 3.3 70B, packet.ai Token Factory undercuts Together AI's $1.04/M rate by 43%.
Key takeaways
DeepSeek built its reputation on one thing: frontier-class model quality at prices that made Western labs uncomfortable. For most of 2026 that meant $0.14 per million input tokens, flat, simple, predictable. That changed three days ago.
On August 16, 2026, DeepSeek activated a peak and off-peak billing structure across all V4 models. For solo builders, indie hackers, and cost-sensitive teams who chose DeepSeek specifically because of the price, this matters. This guide breaks down exactly what changed, what you actually pay now across every major access route, and where packet.ai Token Factory fits into the picture.
For a broader look at LLM inference costs across all major providers, see our LLM inference cost breakdown.
Until August 16, 2026 at 16:00 UTC, DeepSeek V4 Flash billed at a flat $0.14 per million input tokens and $0.28 per million output tokens. Effective that date, DeepSeek switched to a two-tier system based on UTC time of day.
Source: DeepSeek API documentation, accessed August 19, 2026. Always verify at api-docs.deepseek.com before committing budget.
For US-based developers, there is good news buried in the timing. Peak hours run 01:00-04:00 UTC and 06:00-10:00 UTC. That translates to roughly 8pm-1am and 1am-5am US Eastern, outside standard working hours for most solo builders running workloads during the day. If your pipeline runs on a normal US schedule, you are almost always in the off-peak window.
DeepSeek's own API is not the only way to run V4 Flash. Third-party inference providers host the same open weights, often at different prices. Here is every verified route as of August 19, 2026.
Sources: together.ai/pricing (fetched August 19, 2026), openrouter.ai/deepseek/deepseek-v4-flash, api-docs.deepseek.com. Verify before use. Third-party providers may or may not mirror DeepSeek's peak/off-peak schedule.
Together AI at $0.14 input / $0.28 output is currently the cheapest route to DeepSeek V4 Flash 0731, matching the pre-hike DeepSeek direct rate. OpenRouter routes across 18 providers and shows lower blended rates depending on which provider handles your request.
For most solo dev workloads, chatbots, classification, summarisation, RAG retrieval, you do not need DeepSeek V4 specifically. You need a capable open-weight model at the lowest possible cost. packet.ai Token Factory serves several models that undercut Together AI on comparable quality tiers.
Sources: packet.ai pricing data (August 2026), together.ai/pricing (fetched August 19, 2026). Prices per 1M tokens, blended input/output.
packet.ai Token Factory's Llama 3.3 70B at $0.59 per million tokens costs 43% less than Together AI's $1.04 rate for the same model, verified from both providers' official pricing pages on August 19, 2026.
Ready to run inference at these rates? Explore packet.ai Token Factory and start calling models via a simple OpenAI-compatible API.
Per-million-token rates are hard to apply directly. Here is what five common workloads actually cost per month across the cheapest available routes.
Estimates assume a 60/40 input/output split. packet.ai Token Factory pricing from official spreadsheet (August 2026). Together AI pricing from together.ai/pricing (August 19, 2026).
If your volume crosses roughly 5 million output tokens per day, self-hosting DeepSeek V4 open weights on your own GPU starts to beat the API. DeepSeek released V4 Flash under an MIT licence, so the weights are yours to run.
Sources: packet.ai pricing (August 2026), together.ai/pricing (fetched August 19, 2026). All prices per GPU per hour.
packet.ai B200 dynamic at $3.75 per GPU per hour is 54% cheaper than Together AI's on-demand B200 cluster at $8.19 per hour, making it the most cost-effective path for teams running DeepSeek V4 open weights at scale.
To deploy a B200 cluster for self-hosted inference, see packet.ai B200 GPU pricing and availability. For a broader set of GPU options, browse available GPU clusters.
DeepSeek V4 Flash uses an OpenAI-compatible endpoint. If you have an existing OpenAI SDK integration, migrating is a base URL change and a model name update.
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": "user", "content": "Hello"}]
)
print(response.choices[0].message.content)Migration required
The legacy deepseek-chat and deepseek-reasoner model aliases were retired on July 24, 2026. Any code still calling these names will error. Update to deepseek-v4-flash (non-thinking and thinking modes) or deepseek-v4-pro.
DeepSeek applies disk-based prefix caching automatically. No SDK changes or special headers needed. When your prompt shares the same prefix across requests (a system prompt, a retrieved document, tool definitions), the cached portion bills at the cache-hit rate instead of the standard input rate.
At off-peak rates, a cache hit costs $0.007 per million tokens versus $0.22 for a cache miss, a 97% reduction on that portion of your input. For agent loops with a stable instruction set or RAG pipelines with a consistent document prefix, this can cut your effective input cost to near zero.
The practical rule: put stable content (instructions, schemas, documents) at the start of your prompt and keep it byte-identical across calls. Variable content (the user's question) goes at the end.
The right choice depends on your daily volume and what quality floor you need. Here is a direct decision matrix based on verified pricing.
For a deeper look at how self-hosting costs compare to managed inference APIs across different GPU types, see our AI token cost guide.
Last reviewed: August 19, 2026. For the most affordable hosted LLM inference, explore packet.ai Token Factory, or browse GPU clusters if you are ready to self-host open weights.
Same models. Same API. Fraction of the cost. Start free — no credit card required.
Start Building →