Blog Model Guides
DeepSeek V4 Pro and Flash for Production Agents
Use the cheaper candidate for routine steps only when it passes your own acceptance tests.
AIWave prices checked 2026-11-05 UTC against the live pricing response and active default-group routes. The old price and capacity examples on this URL were not current evidence. Check the live rate card before using a number in production.
| Exact model ID | Fresh input | Cache-hit input | Output |
|---|---|---|---|
deepseek-v4-flash | $0.30 | $0.006 | $1.20 |
deepseek-v4-pro | $1.32 | $0.044 | $3.96 |
Route by task evidence
Build a small task set from your actual agent workload: extraction, code edits, tool planning, and review. For each task, freeze the prompt and expected output before running both exact model IDs. Track pass rate, retries, reported token use, and latency. A lower token rate alone cannot establish lower completed-task cost.
Contain the fallback
Pick one model at admission. Escalate a failed task to Pro only under an explicit retry budget, and record both attempts against the task. Do not silently replay non-idempotent tool actions. Keep customer secrets and raw prompts out of public comparison logs.
Forecast the spend
Multiply fresh input, cache-hit input, and output tokens by the matching checked rates, then apply the request's effective account-group multiplier. Use the settled request ledger to confirm the forecast. The table shows default-group base rates; a different group changes the bill.
For an executable first request, use the quickstart. Recheck the effective account group and request settlement before treating an estimate as a bill.