DeepSeek V4 Peak and Off-Peak Routing for Production Agents

The 2026-08-14 AIWave keyword report was generated on 2026-08-15 Asia/Shanghai and could not retrieve current Google Search Console rows because the server credential returned HTTP 401. That failure is important: no search rows should be invented. The usable signal is the report's public market sweep, which points directly at DeepSeek V4 routing, cache economics and dated price checks for Tier 1 and Tier 2 developers. DeepSeek's official pricing page now has two facts that should change any production agent plan: the current V4 Pro and V4 Flash rates are still active before 2026-08-16 16:00 UTC, and a new peak/off-peak table is scheduled to start exactly then.

Keyword source: the 2026-08-14 report recommends refreshing DeepSeek routing content with explicit source dates, cache examples, direct API caveats and a production migration checklist.

Price Windows Checked on 2026-08-15

DeepSeek lists deepseek-v4-flash and deepseek-v4-pro with 1M context, OpenAI-format and Anthropic-format base URLs, JSON output, tool calls, Responses API support, prefix completion, FIM completion in non-thinking mode and separate cache-hit, cache-miss and output rows. The current table before the scheduled change lists Flash at $0.0028 cache-hit input, $0.14 cache-miss input and $0.28 output per 1M tokens. Pro is listed at $0.003625 cache-hit input, $0.435 cache-miss input and $0.87 output per 1M tokens.

That change means an article published today should not talk about DeepSeek V4 as a static price row. It should tell developers exactly which table was checked, which UTC effective time applies, and how their agent scheduler should behave before and after the transition.

Model Split for Agent Workflows

Production agents need two different decisions. The first is capability routing: Flash handles routine work where errors are visible and repairable, while Pro handles reasoning-heavy steps where a bad answer can create operational risk. The second is billing-window routing: after 2026-08-16 16:00 UTC, the same Pro or Flash call will have a different estimate depending on whether it lands in DeepSeek's peak windows. Teams in the United States, United Kingdom, Germany, the Netherlands, Japan and Singapore should store UTC timestamps, not local labels, so finance and engineering can reconcile the route later.

from dataclasses import dataclass
from datetime import datetime, time, timezone
from openai import OpenAI

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

@dataclass(frozen=True)
class WindowPrice:
    cache_hit: float
    cache_miss: float
    output: float
    source_date: str

CURRENT = {
    "deepseek-v4-flash": WindowPrice(0.0028, 0.14, 0.28, "2026-08-15"),
    "deepseek-v4-pro": WindowPrice(0.003625, 0.435, 0.87, "2026-08-15"),
}

AFTER_2026_08_16_1600_UTC = {
    "deepseek-v4-flash": {
        "off_peak": WindowPrice(0.007, 0.22, 0.66, "2026-08-16 16:00 UTC"),
        "peak": WindowPrice(0.014, 0.44, 1.32, "2026-08-16 16:00 UTC"),
    },
    "deepseek-v4-pro": {
        "off_peak": WindowPrice(0.022, 0.66, 1.98, "2026-08-16 16:00 UTC"),
        "peak": WindowPrice(0.044, 1.32, 3.96, "2026-08-16 16:00 UTC"),
    },
}

PEAK_WINDOWS_UTC = [(time(1, 0), time(4, 0)), (time(6, 0), time(10, 0))]

def billing_window(now: datetime) -> str:
    utc = now.astimezone(timezone.utc).time()
    return "peak" if any(start <= utc < end for start, end in PEAK_WINDOWS_UTC) else "off_peak"

def estimate(model: str, input_tokens: int, output_tokens: int, cache_ratio: float, now: datetime) -> dict:
    active = CURRENT if now < datetime(2026, 8, 16, 16, 0, tzinfo=timezone.utc) else {
        k: v[billing_window(now)] for k, v in AFTER_2026_08_16_1600_UTC.items()
    }
    p = active[model]
    cached = input_tokens * cache_ratio
    fresh = input_tokens - cached
    usd = cached / 1_000_000 * p.cache_hit + fresh / 1_000_000 * p.cache_miss + output_tokens / 1_000_000 * p.output
    return {"model": model, "window": billing_window(now), "estimated_usd": round(usd, 6), "source_date": p.source_date}

print(estimate("deepseek-v4-pro", 320_000, 12_000, 0.71, datetime.now(timezone.utc)))

The example deliberately keeps model choice, window choice and estimate outside the prompt. The model should not be asked whether a job is high stakes or whether the current hour is peak. Application code should classify the task, calculate the current billing window and attach a policy version before the request is sent through an OpenAI-compatible client.

Cache and Output Controls

DeepSeek's current and scheduled tables both separate cache-hit input from cache-miss input. Long-context coding agents often resend repository maps, issue histories, retrieval bundles and policy documents. If that shared context is stable, cache logging can materially change the estimate. If the context is repacked on every turn, cache assumptions will not hold. A useful ledger therefore logs shared-context ID, cache-hit ratio, fresh input tokens, output cap, model ID, task class and source date.

• Reserve Pro for security, incident, architecture, migration and root-cause tasks.
• Default to Flash for repeatable execution, extraction, summarization and cleanup steps.
• Attach an output-token cap by task class so output does not dominate the bill.
• Store UTC request time and whether the call used the current or post-change DeepSeek table.
• Keep a rollback route so one workflow can move back to a previous provider path without a deployment.

This is also the SEO angle. Developers searching for DeepSeek API pricing today are not served by stale launch-price snippets. They need a route design that acknowledges the exact effective date and the cache rows, then gives them a way to estimate, cap, observe and reverse production traffic.

Where AIWave Fits

AIWave should use this page to link readers from DeepSeek-specific routing into AIWave Chat Completions, the live models page and pricing review. The platform value is not a single model call. It is a consistent OpenAI-compatible surface for Chinese model families, with the application maintaining policy, budgets, logs and route reasons. This keeps the content aligned with Tier 1 and Tier 2 developer intent rather than broad price-shopping traffic.

Do not claim permanent savings, enterprise adoption numbers or account-specific discounts that were not verified in this run. The credible message is dated source checks, production routing, cache-aware cost estimation, UTC window handling and rollback. That is enough to make the article useful and defensible.

External sources checked

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

Related AIWave guides

FAQ

When does DeepSeek's peak/off-peak pricing start?

DeepSeek's official page says the new peak/off-peak table takes effect at 16:00 UTC on August 16, 2026.

Which DeepSeek V4 model should production agents use first?

Use Flash for routine execution steps and reserve Pro for security, architecture, incident, migration and root-cause reasoning.

Why log UTC time?

UTC timestamps let finance and engineering reconcile whether a request used the current price table or the post-change peak/off-peak table.