DeepSeek V4 Pro and Flash Routing for Production Agents

DeepSeek V4 Pro and Flash make production routing more important than one global model choice. The official DeepSeek pricing snapshot checked on 2026-08-10 lists both models with 1M context, OpenAI and Anthropic format base URLs, JSON output, tool calls, context caching and large output limits. Flash carries lower per-token rates and a larger concurrency limit, while Pro is the escalation path for planning, security, architecture and incident analysis. For developers in the United States, United Kingdom, Canada, Germany, Japan and Singapore, the useful question is how to make that split observable before it reaches a customer invoice.

Keyword source: the 2026-08-10 AIWave keyword report highlights DeepSeek API, DeepSeek access overseas, AIWave API documentation and production routing intent across Tier 1 markets.

Current DeepSeek Price Inputs

Use these values as dated engineering inputs, not permanent commercial terms. DeepSeek says prices can change, so every estimator should carry a checked date and a source URL. The current page separates cached input, fresh input and output tokens; that matters because long-context agents often resend large shared state across many turns.

A basic route policy can start with task class. Flash should handle repeatable work where mistakes are easy to detect or repair. Pro should handle steps where the model is asked to reason across system design, compliance, security, production incidents or multi-stage migration risk. The cache ledger then decides whether a long prompt is economically stable enough to keep in the path.

Why Cache Status Belongs in the Agent Ledger

Token totals alone are not enough. A 210,000-token repository context with a 68 percent cache-hit ratio has a different cost profile from the same context sent as fresh input. Store model ID, task class, input tokens, output tokens, cache-hit ratio, source pricing date, customer region and route reason. That ledger lets engineering and finance review the same evidence when a support bot, coding agent or internal review assistant changes behavior.

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 TokenPrice:
    cached_input: float
    fresh_input: float
    output: float

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

def estimate_usd(model: str, input_tokens: int, output_tokens: int, cache_hit_ratio: float) -> float:
    price = PRICES[model]
    cached_tokens = input_tokens * cache_hit_ratio
    fresh_tokens = input_tokens - cached_tokens
    return round(
        cached_tokens / 1_000_000 * price.cached_input
        + fresh_tokens / 1_000_000 * price.fresh_input
        + output_tokens / 1_000_000 * price.output,
        6,
    )

def route_agent_step(task: str, input_tokens: int, output_tokens: int, cache_hit_ratio: float) -> dict:
    escalation_terms = ("security", "architecture", "incident", "migration", "root cause")
    model = "deepseek-v4-pro" if any(term in task.lower() for term in escalation_terms) else "deepseek-v4-flash"
    return {
        "model": model,
        "estimated_usd": estimate_usd(model, input_tokens, output_tokens, cache_hit_ratio),
        "pricing_checked_at": "2026-08-10",
    }

print(route_agent_step("security migration review", 210_000, 8_000, 0.68))

The preflight example estimates cost before the call is made. In production, the final billing record should come from platform usage, but the preflight estimate gives the application a point to reject a request, cap output tokens, choose Flash, escalate to Pro or ask the user to narrow the context. It also gives Tier 1 teams an audit trail when a customer asks why one task cost more than another.

Production Routing Rules

• Default to Flash for deterministic transformation, extraction, summarization and repeated tool-result cleanup.
• Escalate to Pro for security, architecture, incident, migration and root-cause tasks.
• Keep model selection server-side so a browser client cannot directly choose premium routes.
• Set output-token caps by task class, not by model family alone.
• Run a weekly source check for published price snippets and update the checked date when values change.

AIWave's Chat Completions endpoint is OpenAI-compatible, so a migration can often keep the OpenAI SDK shape while changing the base URL and model ID. That reduces integration friction, but it does not remove the need for policy. A production agent should have route logging, retry limits, output caps and a rollback flag before it handles real customer workloads.

SEO and Conversion Angle

The AIWave report shows that API documentation, pricing and brand queries have Tier 1 impressions but weak click-through. A DeepSeek routing page should therefore avoid generic model hype and answer the operational question directly: how do I split Pro and Flash, estimate cache impact and keep the OpenAI-compatible request path auditable? That is a stronger conversion path than a broad provider ranking because it maps to an engineering task the reader can implement.

Internal links should point to AIWave models, pricing and Chat Completions docs. External links should point to the current DeepSeek pricing page. The article should also make clear that teams must verify live account pricing before rollout, because dated public pages are inputs for planning rather than a billing guarantee.

External sources checked

• https://api-docs.deepseek.com/quick_start/pricing/?article_id=article_1779470751466_8
• https://aiwave.live/docs/chat-completions
• https://aiwave.live/models/
• https://aiwave.live/pricing

Related AIWave guides

FAQ

When should a production agent use DeepSeek V4 Pro?

Use Pro for security, architecture, incident, migration, root-cause and other reasoning-heavy steps where mistakes are expensive.

Why not send every step to Pro?

Flash has lower listed token rates and a larger concurrency signal, so routing routine execution work to Flash can preserve Pro for harder decisions.

What should the ledger store?

Store model ID, task class, input tokens, output tokens, cache-hit ratio, source pricing date, region policy and route reason.