DeepSeek V4 Agent Route Receipts for Pro and Flash

Why This Topic Matters Now

The Sep 18 keyword report highlighted a practical comparison query: DeepSeek versus GLM versus Kimi. For production agents, the useful question is not which model wins globally. It is which route should own planning, execution, summarization, and recovery, and what evidence should move traffic between them.

The current AIWave check gives teams two different surfaces. The live route endpoint exposes 68 rows and current group context. The public pricing JSON exposes dated USD base rates and was updated on Sep 18, while its price rows are checked through Sep 10. That distinction lets a buyer use current route availability without pretending that a dated rate card is a perpetual quote.

Source Facts Checked Today

AIWave /api/pricing was checked from production on Sep 19, 2026 and returned HTTP 200, success=true, 68 live route rows, pricing_version a42d372ccf0b5dd13ecf71203521f9d2, auto_groups=['default'], and group_ratio default=1 and vip=0.9. The public /api/v1/pricing endpoint also returned HTTP 200 with 56 dated USD rows, pricing_version 83f77abde81ee3a096a672ed959ccc096f5d37a45c177ae8e03229456b5415a5, updated_at=2026-09-18, and row checked dates of 2026-09-10. Use the live endpoint for route and group evidence, and the static endpoint for dated public USD rates.

The public pricing JSON checked during this run lists deepseek-v4-pro at $1.914 input, $0.0637362 cache-hit input, and $5.742 output per 1M tokens. It lists deepseek-v4-flash at $0.638 input, $0.0202884 cache-hit input, and $1.914 output. Both rows carry effective_date 2026-08-27 and the endpoint's checked field is 2026-09-10.

The live route response checked on Sep 19 lists deepseek-v4-pro with model_ratio 0.957, completion_ratio 3, cache_ratio 0.0333, and an OpenAI endpoint type. It lists deepseek-v4-flash with model_ratio 0.319, completion_ratio 3, cache_ratio 0.0318, and an OpenAI endpoint type. These are route-governance fields, not a second USD rate card.

DeepSeek's official pricing and quick-start pages were included as provider-source checks in the keyword report. Use those pages for direct-provider context, then use AIWave's dated public JSON and live route response for gateway-specific planning. Do not infer concurrency, data retention, or support guarantees from a model name.

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.

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 openai import OpenAI

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

ROUTES = {
    "plan": {"model": "deepseek-v4-pro", "max_tokens": 1200},
    "execute": {"model": "deepseek-v4-flash", "max_tokens": 650},
}

def run_agent_step(stage: str, prompt: str):
    route = ROUTES[stage]
    response = client.chat.completions.create(
        model=route["model"],
        messages=[{"role": "user", "content": prompt}],
        temperature=0.1,
        max_tokens=route["max_tokens"],
    )
    return {"stage": stage, "model": route["model"], "usage": response.usage,
            "finish_reason": response.choices[0].finish_reason}

Make the Route Decision Explicit

A DeepSeek V4 Pro versus Flash routing decision should name the workload, route, source date, output cap, retry ceiling, data class, budget owner, and fallback. Put those fields in configuration or a review record so a model change is visible to engineering, finance, and support.

Separate Gateway Evidence From Provider Claims

Provider documentation explains capability and direct-platform billing concepts. AIWave endpoints explain the gateway's current route rows, public USD snapshot, supported endpoint type, and key-group context. Keep those evidence surfaces separate; a gateway row is not a promise that every provider feature is exposed through every route.

Use a Small Acceptance Set

Start with redacted, representative prompts instead of the largest production payload. Include one normal case, one long-context case, one malformed-input case, and one stop-condition case. Record finish reason, token usage, route, retry count, reviewer decision, and the checked source date. Promote only after the team can explain both quality and spend.

Budget Input, Cache, and Output Separately

A blended token number hides the decision. Store input tokens, cache-hit input when exposed, output tokens, and tool calls in different columns. Output caps must follow task class: a short router, a code review, and a long planning pass should not inherit one shared allowance.

Keep Retries and Fallbacks Bounded

Retry only errors that the client can safely retry, add an attempt ceiling, and preserve the request identifier across attempts. A fallback should be selected by policy rather than by an unbounded loop. The receipt should show the original route, fallback route, stop reason, and whether the answer was accepted or revised.

Use the AIWave Evidence Layer

Use the Models docs, Chat Completions docs, dated Pricing JSON, Status, and Trust. Read 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 a procurement review. Keep a copy of the checked dates in the internal decision record.

Final Promotion Checklist

Promotion is ready when the provider source is dated, the AIWave route is rechecked, the public USD row is labeled with its own checked date, canaries pass, output and retry limits exist, privacy handling is documented, and a reviewer can reconcile the receipt without seeing a prompt or reusable credential.