This guide uses source checks from Sep 4, 2026. Provider and gateway prices can change; preserve the checked date with every forecast.
Why This Topic Matters Now
The Sep 3 keyword report still points toward DeepSeek V4, DeepSeek pricing, OpenAI-compatible API work, and developer conversations about routing Pro and Flash. The useful question for an engineering team is not which model is universally better. It is whether Pro should plan, Flash should execute, and what evidence proves that split for a specific coding-agent workload.
This guide gives Tier 1 and Tier 2 teams an acceptance-test plan for a DeepSeek worker split. It uses direct DeepSeek price and capability rows checked through the current official docs, plus AIWave gateway rows checked on Sep 4, 2026. The two sources stay separate so the test can support either a direct-provider decision or an AIWave gateway decision without mixing rate cards.
Source Facts Checked Today
DeepSeek official pricing docs crawled for Sep 2026 list deepseek-v4-flash, deepseek-v4-pro, and deepseek-v4-flash-vision-exp with OpenAI and Anthropic base URL formats, 1M context, 384K maximum output, Responses API support, tool calls, JSON output, and cache-hit, cache-miss, and output token rows. The page also lists account-level concurrency limits of 2500 for Flash routes and 500 for Pro.
The same DeepSeek table lists weekday peak windows of 01:00-04:00 and 06:00-10:00 UTC. Flash is shown at $0.007 off-peak and $0.014 peak for cache-hit input, $0.22 off-peak and $0.44 peak for cache-miss input, and $0.66 off-peak and $1.32 peak for output per 1M tokens. Pro is shown at $0.022/$0.044 cache-hit input, $0.66/$1.32 cache-miss input, and $1.98/$3.96 output per 1M tokens.
AIWave /api/pricing checked on Sep 4, 2026 returned success=true, 63 records, pricing_version a42d372ccf0b5dd13ecf71203521f9d2, default group ratio 3, and VIP group ratio 1. Parsed gateway examples before account-group math were V4 Flash at $0.638 input, $1.914 output, and $0.020288 cache-hit input per 1M tokens, and V4 Pro at $1.914 input, $5.742 output, and $0.063736 cache-hit input.
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.
| Acceptance area | Pro worker test | Flash worker test |
|---|---|---|
| Planning | Architecture and risk decomposition | Reject if it skips constraints |
| Execution | Optional patch critique | Patch outline or command plan |
| Cache | Stable repo brief measured | Repeated task preamble measured |
| Output cap | Longer rationale allowed | Short task-specific cap |
| Failure class | Stop on auth or policy mismatch | Retry bounded capacity errors |
| Fallback | Human review before downgrade | Escalate hard tasks to Pro |
| Ledger | Direct and gateway rows separate | Direct and gateway rows separate |
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")
TESTS = [
("planning", "deepseek-v4-pro", "Create a migration risk list from this redacted module map."),
("execution", "deepseek-v4-flash", "Draft the smallest patch plan for this redacted bug report."),
]
for role, model, prompt in TESTS:
response = client.chat.completions.create(
model=model,
messages=[{"role": "user", "content": prompt}],
temperature=0.1,
max_tokens=700 if role == "planning" else 420,
)
print({"role": role, "model": model, "usage": response.usage})
Define Worker Jobs
A Pro and Flash split only makes sense when the jobs are different. Planning workers should decompose risk, identify missing evidence, decide whether a task is safe to automate, and write acceptance criteria. Execution workers should produce small patch plans, test commands, and deterministic transformations. If both workers receive the same prompt and output cap, the split is only a model preference, not an engineering design.
Create a Redacted Fixture Set
Use five fixtures: a dependency upgrade, a schema migration, a failing test, a frontend layout issue, and a long-context repository map. Strip credentials, customer data, and proprietary details that the model does not need. Each fixture should have expected qualities such as correct risk order, no invented files, bounded output, and a clear stop condition. Acceptance tests should fail a model for plausible but ungrounded work, not only for syntax mistakes.
Score Planning Separately
The Pro worker should be judged on constraint recognition, risk order, missing-evidence questions, fallback judgment, and whether it recommends human review at the right time. A concise answer can pass if it protects the system. A long answer should fail if it ignores the data boundary or suggests broad changes without evidence. Store qualitative notes beside token usage so cost and quality can be reviewed together.
Score Execution Separately
The Flash worker should be judged on whether it follows the plan, keeps changes small, names test commands, and refuses to invent missing context. It does not need to write a long architecture memo. Use a tighter max token cap, one retry ceiling, and a direct failure class. If Flash repeatedly escalates because the fixture is ambiguous, that may be a good result: the split is revealing where Pro planning or human review is needed.
Record Cache and Output Assumptions
DeepSeek separates cache-hit and cache-miss input, so repeated repository briefs should be versioned. Put stable system instructions and repo maps first, volatile issue details last, and store prompt_template_version with the ledger. Output caps should also be part of the test. A Pro plan with 900 tokens and a Flash task with 420 tokens are different cost assumptions and should not be averaged into one blended row.
Internal Links for DeepSeek Evaluators
Readers should move through Chat Completions, Models docs, Pricing, Trust, the concurrency runbook, and the 10M-token ledger. That path keeps acceptance testing connected to capacity and spend evidence.
Procurement Review
Procurement should receive a table with fixture name, worker role, model ID, source date, input tokens, cached input tokens, output tokens, retry count, pass or fail, and decision owner. Direct DeepSeek rows and AIWave gateway rows should be adjacent but separate. The review question is not which route has a smaller unit price. It is whether the measured split is good enough for the workload and explainable enough for finance.
Final Checklist
A DeepSeek V4 worker split is ready when job roles are distinct, fixtures are redacted, planning and execution are scored separately, cache assumptions are versioned, output caps are role-specific, retries are classified, and pricing rows are source-dated. If those fields are missing, keep the workflow in canary and run another fixture set before moving real work through the split.