DeepSeek V4 Pro and Flash Routing After the August 16 Price Change

Keyword source: AIWave Daily Keyword Intelligence for 2026-08-16, generated from GSC rows and public market checks for Tier 1 and Tier 2 developer intent.

Keyword Signal and Price Source

The 2026-08-16 keyword report points to a clear Tier 1 problem: developers are still searching for DeepSeek API, DeepSeek V4, AIWave API documentation and OpenAI-compatible Chinese model access, while the United States supplies the largest impression base with weak CTR. That is a content signal, but it is also an engineering signal. A page that only repeats a model name will not help a team in San Francisco, London, Berlin, Tokyo or Singapore plan an actual agent budget.

The publish-time facts are dated. The AIWave keyword report was generated at 2026-08-16 05:12:49 Asia/Shanghai and recorded official DeepSeek pricing rows for V4 Flash and V4 Pro, including cache-hit input, cache-miss input, output, a 1M-token context window and an effective time of 2026-08-16 16:00 UTC for the new table. Public news coverage on 2026-08-14 also reported the output rows that matter most to budgets: V4 Pro at $3.96 per 1M output tokens during peak hours and $1.98 off peak, and V4 Flash at $1.32 peak and $0.66 off peak.

That makes this a routing article, not a simple model comparison. After the effective time, a route record needs to carry four fields that many prototypes skip: UTC request time, model ID, cache hit share and output-token cap. Without those fields, finance cannot explain why a Tuesday batch job cost more than a Sunday job, and engineering cannot tell whether the difference came from peak timing, cache misses, retries or long answers.

AIWave's role is a stable OpenAI-compatible control surface. The application still owns policy: which tasks can use Pro, which can use Flash, which calls can wait for an off-peak window, and which incident or customer-facing paths should run immediately. The content should therefore push readers from the DeepSeek query into a deployable routing pattern, then link them to the AIWave Chat Completions docs, live model catalog and pricing review before production rollout.

Post-Change Price Table

The table below is intentionally narrow. It lists the DeepSeek V4 rows needed by API teams, using USD per 1M tokens and the 2026-08-16 source date. It does not promise a stable future rate, account-specific terms or provider availability. Treat it as a routing input that must be rechecked before a new budget is committed.

Model	Off-peak cache hit	Off-peak cache miss	Off-peak output	Peak cache hit	Peak cache miss	Peak output
deepseek-v4-flash	$0.007	$0.22	$0.66	$0.014	$0.44	$1.32
deepseek-v4-pro	$0.022	$0.66	$1.98	$0.044	$1.32	$3.96

The engineering takeaway is not simply that Flash costs less than Pro. The more useful point is that output now dominates many V4 Pro sessions. A root-cause analysis prompt with 300,000 input tokens and a 20,000-token response can stay manageable if most input is cached, but the output side still needs a cap and a review policy. A coding agent that lets every step produce long traces can create a cost pattern that looks like a traffic spike even when request count is steady.

Cache rows also need care. A cache-hit row only applies to tokens the provider reports as hits. Repacking the same repository map with a different timestamp, random file order or new tool schema can turn a planned cache hit into a cache miss. Route code should use stable prefixes, stable tool descriptions and explicit measurement rather than assuming that repeated business logic automatically qualifies for the lower input row.

Routing Policy for Agents

A practical policy starts with task class, not model brand. Pro is justified when a wrong answer has high operational cost: migration strategy, security triage, incident analysis, complex code review, financial reconciliation and architecture decisions. Flash is the default for extraction, classification, short summaries, codebase cleanup, tool-result compression and follow-up checks where errors can be detected cheaply.

The second layer is time. If a task is interactive or tied to an incident, run it immediately and log the peak/off-peak status. If it is an async evaluation, nightly doc refresh, benchmark sweep or non-urgent refactor analysis, schedule it outside peak windows when the business can tolerate delay. The important part is to record why the scheduler waited. A silent delay looks like latency; a logged policy decision looks like cost governance.

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://aiwave.live/v1",
)

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

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

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

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

def route_model(task_class: str, urgency: str) -> str:
    high_risk = {"security_review", "incident_rca", "architecture", "migration_plan"}
    if task_class in high_risk or urgency == "critical":
        return "deepseek-v4-pro"
    return "deepseek-v4-flash"

def estimate_usd(model: str, input_tokens: int, output_tokens: int, cache_hit_ratio: float, now: datetime) -> dict:
    row = PRICES[model][billing_window(now)]
    cached = input_tokens * cache_hit_ratio
    fresh = input_tokens - cached
    total = cached / 1_000_000 * row.cache_hit
    total += fresh / 1_000_000 * row.cache_miss
    total += output_tokens / 1_000_000 * row.output
    return {"model": model, "window": billing_window(now), "estimated_usd": round(total, 6), "source_date": row.source_date}

print(estimate_usd("deepseek-v4-pro", 300_000, 18_000, 0.72, datetime.now(timezone.utc)))

The sample keeps the price table in application code rather than inside the prompt. That distinction matters. The model can explain a route, but the application should own the policy, the model ID, the current window and the budget estimate. It also keeps `YOUR_API_KEY_HERE` as a placeholder so no real credential leaks into documentation, test output or source control.

Operational Checklist

Before switching a real agent to the post-change V4 table, run a one-day shadow ledger. For every request, record provider usage fields, route reason, task class, end-user workflow, UTC time, cache-hit tokens, cache-miss tokens, output tokens, retries and final status. Compare the ledger against invoice rows after billing catches up. If the two diverge, fix the measurement before increasing traffic.

Use output caps by task class. A policy might allow 2,000 output tokens for extraction, 6,000 for code review, 12,000 for migration planning and a manual override for incident response. The cap is not only a budget guard. It also makes behavior easier to evaluate because excessively long answers no longer hide quality problems behind verbosity.

Keep a rollback path. One OpenAI-compatible endpoint lets teams move a route from V4 Pro to V4 Flash, GLM, Kimi, Qwen or another approved model without rewriting client code. The application still needs a feature flag, a canary cohort and a route ledger. Otherwise a rollback becomes a manual emergency edit.

Finally, link the DeepSeek article to the AIWave docs and models page. The keyword report says AIWave API documentation already receives Tier 1 impressions with weak clicks. Internal links from model-specific content to `/docs/chat-completions`, `/models/` and `/pricing` are a direct CTR and conversion cleanup step, not decoration.

External sources checked

Related AIWave guides

FAQ

What changed for DeepSeek V4 on August 16, 2026?

The checked pricing data says DeepSeek V4 Pro and Flash moved to peak and off-peak USD rows effective 2026-08-16 16:00 UTC, so production ledgers need a source date and UTC request time.

Should every agent step use DeepSeek V4 Pro?

No. Keep Pro for planning, architecture, security review and root-cause analysis, then use Flash for execution loops, extraction and repeatable cleanup steps.

What should teams log for DeepSeek cache economics?

Log cache-hit input tokens, cache-miss input tokens, output tokens, model ID, route reason, UTC window and the price-source date.