This guide uses source checks from Sep 21, 2026. Provider and gateway prices can change; preserve the checked date with every forecast.
Why This Topic Matters Now
The Sep 20 keyword report proposed DeepSeek V4 Pro versus Flash routing for production agents because the practical question is not which model wins a leaderboard. It is which step deserves deeper reasoning, which step needs predictable throughput, and how a team can explain the choice after the request is complete.
A useful split is planning versus execution. Pro can be the candidate for decomposition, high-impact decisions, and final review. Flash can be the candidate for bounded extraction, tool argument drafting, and repetitive execution when a fixed acceptance set shows it is sufficient. This is a policy pattern, not a claim that every workload should use the same model.
Source Facts Checked Today
AIWave /api/pricing was checked from production on Sep 21, 2026 and returned HTTP 200, success=true, 68 live route rows, pricing_version 5a90f2b86c08bd983a9a2e6d66c255f4eaef9c4bc934386d2b6ae84ef0ff1f1f, auto_groups=['default'], group_ratio default=1 and vip=0.9, with the OpenAI-compatible POST path at /v1/chat/completions. The public /api/v1/pricing endpoint returned HTTP 200 with 56 dated USD rows, pricing_version 83f77abde81ee3a096a672ed959ccc096f5d37a45c177ae8e03229456b5415a5, updated_at=2026-09-18, and checked=2026-09-10. Use the live endpoint for route and group evidence, and the static endpoint for dated public USD rates.
The dated public USD rows checked in this run list deepseek-v4-pro at $1.914 input, $0.0637362 cache-hit input, and $5.742 output per 1M tokens, and deepseek-v4-flash at $0.638 input, $0.0202884 cache-hit input, and $1.914 output per 1M tokens. The rows carry effective_date 2026-08-27, while the response carries checked=2026-09-10 and updated_at=2026-09-18.
The live route response includes deepseek-v4-pro and deepseek-v4-flash with the OpenAI-compatible endpoint type. That confirms route evidence for this check; it does not prove provider-side retention, regional processing, latency, or quality for a particular agent.
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.
| Agent step | Candidate policy | Acceptance evidence |
|---|---|---|
| Task decomposition | Pro first | constraint coverage and reviewer score |
| Tool argument draft | Flash first | schema validity and dry-run result |
| Large repeated context | measure cache share | prefix version and token fields |
| High-impact decision | Pro with human gate | decision record and stop condition |
| Routine extraction | Flash with output cap | field accuracy and finish reason |
| Fallback | one bounded alternate | attempt count and route receipt |
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 dataclasses import dataclass
from openai import OpenAI
client = OpenAI(api_key="YOUR_API_KEY_HERE", base_url="https://aiwave.live/v1")
@dataclass
class RoutePolicy:
planning_model: str = "deepseek-v4-pro"
execution_model: str = "deepseek-v4-flash"
max_attempts: int = 2
def choose_model(step: str) -> str:
return RoutePolicy().planning_model if step in {"plan", "review"} else RoutePolicy().execution_model
def run(step: str, prompt: str):
return client.chat.completions.create(
model=choose_model(step),
messages=[{"role": "user", "content": prompt}],
temperature=0.0, max_tokens=500,
)
Turn the Search Intent Into a Runbook
The useful unit for a DeepSeek V4 Pro and V4 Flash agent router is a runbook, not a model slogan. Write down the workload, approved model IDs, source dates, data class, output ceiling, retry ceiling, owner, and stop condition before the first production request. That record gives engineering, finance, and privacy reviewers the same object to inspect.
Separate Live Route Evidence From Dated Prices
The current route response and the public USD snapshot answer different questions. The live response tells you which route rows and endpoint types are available now. The public snapshot gives dated base rates for a forecast. Keep both URLs, versions, checked dates, and model IDs in the release record instead of blending them into one timeless table.
Measure the Workload You Actually Ship
A short demo can hide the important cost and reliability behavior. Build an acceptance set with ordinary input, repeated context, a long document, a malformed request, and a stop-condition case. Capture input tokens, cached input when exposed, output tokens, tool calls, retries, finish reason, request identifier, and reviewer outcome.
Protect the Request Boundary
Keep credentials server side, use a placeholder in documentation, redact customer content from test fixtures, and make route policy explicit in configuration. OpenAI compatibility reduces client changes; it does not decide what data may cross a route or what a reviewer must retain.
Use Bounded Recovery
Retry only errors that are safe to retry. Put an attempt ceiling on every fallback and preserve the original request identifier. A receipt should show the original route, fallback route, stop reason, and whether the output was accepted, revised, or discarded.
Use AIWave's Evidence Layer
Use the Models docs, dated Pricing JSON, live route pricing, and Status. Recheck the live route table before a rollout, the dated pricing JSON before a budget review, the status page before a launch window, and the trust page before procurement review. Keep source dates visible in the internal decision record.
Final Release Gate
Promotion is ready when provider sources are dated, AIWave routes are rechecked, public USD rows carry their own checked dates, the representative canary passes, data handling is documented, and a named owner can stop or reverse the change.