This guide uses source checks from Aug 28, 2026. Provider and gateway prices can change; preserve the checked date with every forecast.
Why This Topic Matters Now
The Aug 28 publication run did not find a newer keyword report, so this article uses the latest available keyword intelligence from server-date 2026-08-26 and the topic queue built on Aug 27. That queue puts one question first: what does a large DeepSeek V4 Pro workload really cost from outside China? The useful answer is not a slogan or a single blended rate. It is a ledger that separates input, cache-hit input, output, retries, and billing group.
For Tier 1 teams in the United States, United Kingdom, Canada, Germany, Netherlands, Japan, and Singapore, the practical buying motion is often a migration estimate. A team has a RAG system, coding agent, or document workflow that might send millions of tokens through a Chinese model route. The buyer needs enough math to approve a controlled trial and enough telemetry to explain the first invoice. This guide shows the exact columns to collect before production traffic moves.
Source Facts Checked Today
AIWave /api/pricing checked on Aug 28, 2026 returned success, 63 model records, 9 provider families, group_ratio values of default 3 and vip 1, and pricing_version 5a90f2b86c08bd983a9a2e6d66c255f4eaef9c4bc934386d2b6ae84ef0ff1f1f. The live row for deepseek-v4-pro translates to $1.914 per 1M input tokens, $5.742 per 1M output tokens, and $0.063736 per 1M cache-hit input tokens. The row is enabled for default, vip, and svip groups.
The same live feed listed deepseek-v4-flash at $0.638 input, $1.914 output, and $0.020288 cache-hit input per 1M tokens. That matters because many high-input systems use Pro for planning, review, or difficult reasoning, then use Flash for routine execution. A 10M-token worksheet should therefore include an optional split-route tab instead of assuming every request must use the same route.
DeepSeek official pricing checked on Aug 28, 2026 remains the direct-provider context for V4 Flash and V4 Pro, including cache-hit, cache-miss, output, peak and off-peak rows, 1M context, 384K maximum output, and published concurrency limits. Keep that direct-provider table separate from AIWave's gateway table. A buyer should see which source owns each number, which date it was checked, and whether the estimate uses AIWave all-day rows or direct peak-window rows.
Planning Matrix
A source-dated planning matrix keeps the page useful for engineers and procurement reviewers. It turns a search query into an auditable route decision instead of a loose model preference.
| Ledger field | Example value | Why it belongs in the estimate |
|---|---|---|
| Source owner | AIWave gateway or direct DeepSeek | Prevents row mixing |
| Checked date | 2026-08-28 | Makes future price drift reviewable |
| Model route | deepseek-v4-pro | Links usage to a billable route |
| Input tokens | 10,000,000 | Sets the high-input workload size |
| Cache-hit input | Measured, not guessed | Separates repeated context |
| Output tokens | Scenario-specific | Often drives variance |
| Group multiplier | default 3 or vip 1 | Explains account-level billing |
Implementation Pattern
The implementation pattern keeps credentials as placeholders, pins the AIWave base URL, records the model, and leaves room for route-specific controls. Production applications should move credentials into environment or secret storage.
from decimal import Decimal
AIWAVE_API_KEY = "YOUR_API_KEY_HERE"
PRO = {
"input_per_1m": Decimal("1.914"),
"output_per_1m": Decimal("5.742"),
"cache_hit_per_1m": Decimal("0.063736"),
}
def estimate_deepseek_pro(input_tokens, output_tokens, cache_hit_tokens=0, group_multiplier=1):
uncached = Decimal(input_tokens - cache_hit_tokens) / Decimal(1_000_000)
cached = Decimal(cache_hit_tokens) / Decimal(1_000_000)
output = Decimal(output_tokens) / Decimal(1_000_000)
subtotal = uncached * PRO["input_per_1m"] + cached * PRO["cache_hit_per_1m"] + output * PRO["output_per_1m"]
return round(subtotal * Decimal(group_multiplier), 4)
print(estimate_deepseek_pro(10_000_000, 1_200_000, cache_hit_tokens=4_000_000, group_multiplier=1))
Start With a 10M Input Baseline
A 10M-token worksheet is large enough to reveal the economics of long prompts without pretending to represent every workload. Put the baseline in one row: 10,000,000 input tokens, a measured cache-hit field, a separate output field, route name, billing group, source URL, and checked date. From there, create scenarios for repeated context, long answers, and split routing. The finance team can then see whether variance comes from traffic, cache behavior, output length, or account group rather than a vague model change.
Keep Cache-Hit Tokens Explicit
DeepSeek and AIWave both expose the idea that repeated context should not be priced like fresh input when a cache-hit row applies. The ledger should not replace that field with a percentage in prose. Store the raw cache-hit token count, the raw uncached token count, and the prompt-template version. If a system prompt, policy pack, or repository summary changes every deployment, the cache share can move sharply. The forecast should treat cache behavior as measured telemetry, not a permanent constant.
Model Output Separately
High-input workloads can still become output-heavy when agents explain decisions, draft patches, produce citations, or retry long plans. A 10M input estimate should include at least three output scenarios: restrained summaries, normal agent responses, and long planning output. Set route-specific max tokens before the first trial. If the team later raises an output cap, record the reason and the owner. That one habit keeps the ledger useful when a monthly bill changes.
Apply Group Multipliers Deliberately
The live pricing feed exposes group_ratio values, so the worksheet needs a visible group column. A default account and a vip account can produce different effective costs for the same route if the billing multiplier differs. Do not promise that a top-up or account action automatically changes the group. Instead, verify the effective group in the account, record it beside the request sample, and keep group changes as dated procurement events.
Use Split Routing for Real Agents
A DeepSeek V4 Pro estimate should not force every call through Pro. Many agents can use Pro for architecture, hard debugging, or policy-sensitive review while sending routine transformation to Flash. The worksheet should therefore include a route_mix field: Pro input, Pro output, Flash input, Flash output, and shared cache assumptions. If quality tests show Flash can handle the routine path, the estimate becomes more accurate and the production design becomes easier to govern.
Internal Links for the Buyer Journey
Send readers to AIWave Pricing, Predictable Pricing, Models docs, Chat Completions docs, Trust, and the earlier DeepSeek reasoning-route guide. This article covers the high-input ledger; the older article covers route selection.
Procurement Review
Procurement should ask for six items before approving a large DeepSeek Pro trial: the AIWave price row, the direct DeepSeek context row, the checked dates, the effective account group, the expected cache share, and the output cap. Engineering should attach one redacted request log with model, usage, latency, status, and route owner. That evidence is enough to approve a bounded experiment without turning the discussion into a model-brand debate.
Final Checklist
Before routing high-input traffic, create the ledger, verify the live price row, record the effective group, cap output, split cache-hit from uncached input, and test whether Flash can handle routine work. Recheck direct-provider pricing only as context and do not merge it into AIWave's row. The result is a 10M-token estimate that a Tier 1 buyer can audit later, even if the provider market changes.