This guide uses source checks from Sep 8, 2026. Provider and gateway prices can change; preserve the checked date with every forecast.
Why This Topic Matters Now
DeepSeek V4 Pro and DeepSeek V4 Flash are easy to discuss as a planning model and an execution model. They are harder to run responsibly inside a production agent queue. A queue turns one model choice into many small decisions: which worker may call Pro, which worker may call Flash, which prompt classes can reuse cached context, which requests may retry, and which failure should stop the job instead of silently switching behavior. The Sep 7 keyword report keeps DeepSeek V4 pricing and routing in the Tier 1 content set, so today's article narrows the angle to receipts rather than another broad comparison.
A route receipt is a compact record that an engineering lead, finance reviewer, and security reviewer can all read. It should show the model string, source pricing row, checked date, account group, input tokens, cache-hit input, output tokens, queue role, retry count, and fallback policy. That record matters for Tier 1 and Tier 2 teams because a long coding agent can spend more through repeated context and retries than through a single visible answer. The useful question is not only which DeepSeek route is stronger. It is whether the team can explain why each route was called.
Source Facts Checked Today
AIWave /api/pricing checked on Sep 8, 2026 returned success=true, 63 records, pricing_version a42d372ccf0b5dd13ecf71203521f9d2, default group ratio 3, and VIP group ratio 1. Computed public base examples per 1M text-token units were DeepSeek V4 Flash at $0.638 input, $0.020288 cache-hit input, and $1.914 output; DeepSeek V4 Pro at $1.914 input, $0.063736 cache-hit input, and $5.742 output; GLM-5 at $1.55 input, $0.400001 cache-hit input, and $4.96 output; GLM-5.1 at $2.10 input, $0.680001 cache-hit input, and $6.60 output; Kimi K3 at $4.50 input, $0.90 cache-hit input, and $22.50 output; qwen3.6-flash at $0.267805 input and $1.606832 output.
DeepSeek official pricing docs checked on Sep 8, 2026 returned HTTP 200 and listed V4 Flash, V4 Pro, and V4 Flash Vision Exp with OpenAI and Anthropic base URL compatibility, 1M context, 384K maximum output, cache-hit, cache-miss, output, peak and off-peak rows, and account-level concurrency caps. The keyword report recorded Flash route caps at 2500 and Pro route caps at 500. Those direct-provider facts are useful context, but they are not the same row owner as AIWave gateway pricing.
Public developer analyses from S5 Labs and AIREITER were retained only as market context for why source dates and peak-window caveats matter. The publishable article uses AIWave public rows for gateway examples and official DeepSeek docs for direct-provider context. It avoids broad operating and customer-scale claims because D-013 and D-014 keep public proof limited to capability facts, dated rates, route evidence, and trust boundaries.
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.
| Queue decision | Receipt field | Control |
|---|---|---|
| Planning worker uses Pro | model=deepseek-v4-pro | Require explicit task class and output cap |
| Execution worker uses Flash | model=deepseek-v4-flash | Allow only approved low-risk prompt classes |
| Repeated repository context | cache-hit input field | Version stable prompt prefix and measure hit share |
| Retry after timeout | retry_count and stop_reason | Set retry ceiling before launch |
| Fallback after 429 | fallback_policy | Do not switch model family without acceptance |
| Budget review | source_checked_at and group | Keep AIWave rows separate from direct DeepSeek rows |
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
from datetime import datetime, timezone
client = OpenAI(api_key="YOUR_API_KEY_HERE", base_url="https://aiwave.live/v1")
ROUTES = {
"planner": {"model": "deepseek-v4-pro", "max_tokens": 900},
"executor": {"model": "deepseek-v4-flash", "max_tokens": 550},
}
def run_agent_step(role: str, prompt: str) -> dict:
route = ROUTES[role]
started = datetime.now(timezone.utc).isoformat()
response = client.chat.completions.create(
model=route["model"],
messages=[{"role": "user", "content": prompt}],
temperature=0.1,
max_tokens=route["max_tokens"],
)
return {
"role": role,
"model": route["model"],
"source_checked_at": "2026-09-08",
"pricing_version": "a42d372ccf0b5dd13ecf71203521f9d2",
"request_started_at": started,
"usage": response.usage,
}
print(run_agent_step("executor", "Summarize this redacted patch queue item."))
Start With Queue Roles
Define queue roles before selecting models. A planner may compare approaches, write risk notes, or explain a migration path. An executor may produce a small patch, rewrite tests, or classify an error. A reviewer may only check whether the answer follows the policy. Once roles are named, route selection becomes an explicit configuration decision rather than a prompt habit. That makes the later receipt useful because every call can be tied to the work it was allowed to perform.
Pin the Source Row
Every receipt should store the AIWave pricing URL, DeepSeek official pricing URL, checked date, pricing_version, and account group. The source row is not trivia. It is the only way to explain why a forecast made on Sep 8 differs from a forecast made later. Keep input, cache-hit input, and output in separate columns. If the account used the default group, preserve that multiplier in the workbook instead of presenting a base row as the effective bill.
Measure Cache Before Forecasting
Agent queues often repeat a system prompt, repository summary, policy block, and schema instructions. That creates a reason to measure cache-hit input, but not a reason to assume it. Store the prompt-template version and the stable-prefix byte range beside the usage record. If a worker rewrites the prefix on every retry, a cache-based forecast will be too optimistic. The receipt should expose that quickly enough for finance and engineering to correct the queue.
Limit Output by Role
Output can dominate a DeepSeek V4 Pro or Flash bill when agents write explanations, patches, test plans, and alternatives in one response. Use lower output caps for execution and higher caps only for planning or review tasks that need them. Capture the cap in the receipt. If a task repeatedly hits the cap, split the workflow into a planning call and a patch call instead of raising every worker to a larger allowance.
Handle Capacity Without Silent Drift
The official DeepSeek table exposes different concurrency context for Flash and Pro routes. Whether a team calls DeepSeek directly or through a gateway, capacity planning should not be hidden inside retry code. A receipt should record retries, timeout class, 429 class, and fallback rule. If fallback changes behavior, the job should stop for review or move only to an approved paired route. Silent fallback can make a successful run harder to trust.
Use AIWave Links for Verification
A DeepSeek evaluator should move through Pricing, Status, Chat Completions, Models docs, Trust, and the worker split acceptance guide. Those links give pricing, current probes, request shape, model names, and trust boundaries without relying on private operating metrics.
Procurement Review
A procurement packet should include one receipt from the planner route and one from the executor route. Each record should contain model string, source URLs, checked date, pricing_version, account group, token fields, prompt data class, queue role, output cap, retry policy, and fallback policy. The packet should label AIWave gateway rows separately from direct DeepSeek rows. That separation keeps the review defensible when either page changes.
Final Checklist
DeepSeek V4 Pro and Flash are ready for a small agent-queue trial when roles are named, route strings are pinned, redacted prompts are used, token classes are captured, output caps are role-specific, cache-hit assumptions are measured, retries are bounded, and fallback cannot silently change behavior. Recheck AIWave pricing and official DeepSeek pricing before production traffic or monthly budget approval.