Back to blog
This site is an independent third-party technical services platform offering aggregated access to multiple model APIs. It is not affiliated with, authorized by, or partnered with Anthropic, OpenAI, Google, or any other model provider.
Key Takeaways
- DeepSeek announced new pricing on August 13. The changes take effect at 00:00 Beijing Time on August 17, 2026. Peak hours are 9:00–12:00 and 14:00–18:00; off-peak prices are half the peak rates.
- During peak hours, DeepSeek-V4 Pro costs ¥0.3 / ¥9 / ¥27 per million tokens for cached input / uncached input / output. V4 Flash costs ¥0.10 / ¥3 / ¥9. Off-peak rates are half these prices.
- The sharpest increase applies to cached input, weakening architectures that previously relied on extremely cheap caching.
- OpenCode Go has also reduced its DeepSeek V4 Flash quota allocation. Community reports suggest the allocation fell from the equivalent of the $60 tier to the $15 tier after a previous double-quota promotion ended. Verify the exact details in its documentation and console.
- The practical conclusion for developers is simple: the unit-cost model for agent and batch workloads needs to be rebuilt. There is no need to switch in panic, but do not use the old formula to estimate next month’s bill.
A Closer Look
1. Peak/Off-Peak Pricing Is a Structural Change
DeepSeek’s pricing has evolved from nighttime discounts in February 2025, to long-term discounts for V4 Pro in May 2026, and now to formal peak/off-peak pricing across the product line.
The important point is not simply how many times more expensive the service has become. It is that the relative prices of the three token categories have been rearranged. The approximate ratio of cached input : input : output is now 1 : 30 : 90.
This affects two types of projects very differently:
- Long-context agents with high cache-hit rates: These projects previously stored large system prompts and codebase context in the cache, making marginal costs almost negligible. Cached-input pricing has increased the most, so the economics of this architecture have weakened considerably.
- Short-request, output-heavy conversational projects: Output prices have also increased, but if overall usage is small, the absolute increase may remain limited.
Some developers estimate that, under a typical agentic-coding token mix, the blended price has risen to slightly more than three times the previous level for Pro and more than twice the previous level for Flash. During peak hours, the cost doubles again. This is large enough to change a product’s gross-margin model.
2. The Service Provider Controls the Meaning of Plan Quotas
OpenCode Go defines quotas in terms of “dollar value.” You receive a credit pool, and the number of requests it supports depends on model pricing. When the model price changes, the number of requests supported by the same quota changes as well.
That is what happened with V4 Flash. It is a natural consequence of the pricing mechanism rather than an isolated surprise.
For users, however, the impact is very real. Based on community calculations, the number of V4 Flash requests supported by the same subscription has fallen by nearly 90%. Batch workloads planned during the previous double-quota promotion may need to be rescheduled.
The broader lesson is this: in any plan-based credit system, the conversion between quota and model price is defined by the service provider and can change when upstream pricing changes.
This is not automatically bad. Users with large, stable workloads may still benefit from volume pricing. But plan-based credits should not be your only cost foundation, especially when usage fluctuates or you may not consume the full quota.
What This Means for Developers
1. Calculate before deciding whether to switch.
Export the last 30 days of request logs. Track cached input, uncached input, and output tokens separately, then recalculate the bill using the new price table. If output represents only a small share of usage, the actual increase may be less than double—even if the headline increase sounds like three times.
2. Move flexible workloads to off-peak hours.
Data cleaning, batch labeling, offline evaluation, and document summarization can usually run during off-peak windows, saving 50%. Adding an “only dequeue during off-peak hours” switch to your task queue is relatively inexpensive.
3. Reevaluate caching.
Caching is still useful; it is simply no longer “almost free.” Check whether you are adding redundant context solely to achieve cache hits. What was previously harmless now has a real cost.
4. Do not rely on a single model or billing model.
When an upstream provider changes its prices, downstream plans that convert quotas based on those prices are affected as well. Allowing your business to switch smoothly between multiple models is itself a form of cost hedging.
Multi-Model Routing Through Code0
Code0 is a multi-model API gateway that provides access to more than 300 models, including Claude, GPT, Gemini, DeepSeek, Kimi, Qwen, and Grok.
It is compatible with the OpenAI SDK, and switching models requires changing only the model field. Your business code does not need to change:
from openai import OpenAI
client = OpenAI(
base_url="https://hk.code0.ai/v1",
api_key="sk-your-key", # Get your key from console.code0.ai
)
# Route by task difficulty: change only the model field
ROUTE = {
"light": "deepseek-v4-flash", # Classification, extraction, formatting
"regular": "claude-sonnet-4-6", # Everyday chat, code completion
"hard": "claude-opus-4-8", # Complex reasoning, long-horizon agents
}
def ask(level: str, prompt: str) -> str:
resp = client.chat.completions.create(
model=ROUTE[level],
messages=[{"role": "user", "content": prompt}],
)
return resp.choices[0].message.content
print(ask("light", "Extract the error codes from these logs as JSON"))
print(ask("hard", "Analyze the concurrency flaws in this module and propose fixes"))
Switching to gpt-5.4 or gemini-3-pro requires the same one-line change. To run an A/B comparison, send the same prompt to two models and compare the results—there is no need to maintain separate SDKs or authentication systems.
Code0 uses pay-as-you-go billing and does not charge for failed requests. There is no need to pre-purchase plan credits, so quota-conversion ratios are not subject to this type of adjustment.
Exact model pricing is shown in the Code0 console. Actual savings vary according to your workload’s token structure, so it is best to run a small one-week comparison using real traffic.
Conclusion
Viewed together, the pricing and quota changes send a consistent signal: China’s foundation-model industry is moving away from aggressive price wars and toward pricing frameworks that more closely reflect compute costs.
Low prices will not remain the default forever. That is not necessarily bad for the ecosystem, but it does mean the period of “adjusting things without doing the math” is over.
In the short term, schedule flexible work during off-peak hours and recalculate the value of caching. In the medium term, ensure that your system is not locked to any single model or billing model.
