DeepSeek API Cache Math for Production Agents

DeepSeek API pricing changed the routing question from which model is stronger to which model should handle each step of a production workflow. On 2026-08-09, the official DeepSeek pricing page listed both deepseek-v4-flash and deepseek-v4-pro with 1M context, OpenAI-format and Anthropic-format base URLs, JSON output, tool calls, context caching and distinct concurrency limits. For a US, UK, Canada, Germany, Japan or Singapore SaaS team, the important move is to log cache status, output tokens and task type before the request leaves the application.

Keyword source: the 2026-08-09 AIWave keyword report could not refresh GSC because outbound GSC access was blocked, so this article uses the last verified Tier 1 context from 2026-08-08: aiwave api, aiwave pricing, aiwave api documentation, deepseek api access overseas and DeepSeek comparison queries.

Current DeepSeek Price Inputs

The official DeepSeek table, checked on 2026-08-09, prices each 1M tokens separately by cache-hit input, cache-miss input and output. It also states that product prices may change, so the checked date belongs in every estimator and customer-facing calculator. The point is not to promise a permanent rate; it is to make every routing decision reproducible.

The same page lists concurrency limits of 2500 for Flash and 500 for Pro. Treat that as a capacity planning signal, not just a pricing footnote. A route that sends every task to Pro may look acceptable in a spreadsheet but fail under parallel agent traffic. A route that never escalates can reduce answer quality on the exact steps that customers remember.

Why Cache Math Belongs in the Ledger

A cache-aware request ledger should store prompt tokens, output tokens, cache-hit ratio, model, route reason and the pricing table date. Without those fields, a team cannot explain why yesterday's repository analysis cost less than today's incident review. Cache economics are especially visible on long prompts: a 180,000-token codebase context with a 74 percent cache-hit ratio is a different product experience from a one-off 180,000-token prompt.

from dataclasses import dataclass
from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY_HERE", base_url="https://api.aiwave.live/v1")

@dataclass(frozen=True)
class Price:
    cache_hit_input: float
    cache_miss_input: float
    output: float

PRICES = {
    "deepseek-v4-flash": Price(0.0028, 0.14, 0.28),
    "deepseek-v4-pro": Price(0.003625, 0.435, 0.87),
}

def estimate(model: str, prompt_tokens: int, output_tokens: int, cache_hit_ratio: float) -> float:
    price = PRICES[model]
    cached = prompt_tokens * cache_hit_ratio
    uncached = prompt_tokens - cached
    return round(cached / 1_000_000 * price.cache_hit_input + uncached / 1_000_000 * price.cache_miss_input + output_tokens / 1_000_000 * price.output, 6)

def choose_model(task: str, prompt_tokens: int, expected_output_tokens: int, cache_hit_ratio: float) -> dict:
    complex_terms = ("architecture", "incident", "security", "migration", "review")
    model = "deepseek-v4-pro" if any(term in task.lower() for term in complex_terms) else "deepseek-v4-flash"
    return {"model": model, "estimated_usd": estimate(model, prompt_tokens, expected_output_tokens, cache_hit_ratio), "pricing_checked_at": "2026-08-09"}

print(choose_model("migration review for a Canada SaaS API", 180_000, 6_000, 0.74))

The example does not call the API; it makes the route explain itself before the application sends the request. That preflight estimate is where product teams can apply per-customer budgets, region policy, retry policy and escalation rules. The final invoice should still come from provider or platform usage records, but the estimate keeps the application from flying blind.

A Practical Routing Pattern

Start with Flash for deterministic, repeatable work: classification, extraction, short transformations, retrieval compression, conversation summaries and tool-result cleanup. Escalate to Pro when the user asks for architecture, incident response, security tradeoffs, multi-file reasoning or a migration plan. When the same customer context is reused across turns, record the cache-hit ratio and prefer a stable prompt prefix so cache behavior has a chance to become predictable.

• Keep model choice in server-side policy, not in browser code.
• Store the exact price table date next to each estimate.
• Separate retries from fresh user requests in billing analytics.
• Alert on output-token spikes, because output dominates many reasoning-heavy tasks.
• Run weekly source-page checks before publishing pricing snippets in docs or sales material.

AIWave's public docs show an OpenAI-compatible Chat Completions endpoint, so teams migrating existing OpenAI SDK code can keep the client shape while changing the base URL and model IDs. That makes DeepSeek routing a policy problem rather than a full client rewrite.

Tier 1 SEO and Conversion Angle

The last verified GSC context showed Tier 1 impressions with weak CTR around docs, pricing and API queries. A useful DeepSeek article should therefore answer the operational question directly: how do I estimate a request, pick Pro versus Flash and keep a bill explainable to engineering leadership? Avoid vague cost claims. Developers in the United States, United Kingdom, Germany and Japan need a dated table, runnable code and a migration path.

The best internal links for this topic are AIWave models, pricing and Chat Completions docs. The article should also link to the official DeepSeek pricing page so readers can re-check the rate card before deployment. That combination builds trust without relying on unverifiable benchmark or customer claims.

External sources checked

• https://api-docs.deepseek.com/quick_start/pricing/
• https://aiwave.live/models/
• https://aiwave.live/docs/chat-completions

Related AIWave guides

FAQ

Which DeepSeek model should a production agent use first?

Use deepseek-v4-flash for repeatable execution work and escalate to deepseek-v4-pro for planning, incident analysis, security review or difficult migration steps.

Why should cache-hit ratio be stored in the request ledger?

Because cache-hit and cache-miss input tokens have different prices, and repeated long-context prompts can change the cost profile of the same workflow.

Can an OpenAI SDK app call DeepSeek through AIWave?

Yes. AIWave documents an OpenAI-compatible Chat Completions endpoint, so migration can usually keep the SDK shape while changing the base URL, API key and model ID.