DeepSeek V4.1 Flash is now the current efficiency-focused API option, but V4 Pro has not disappeared immediately. DeepSeek’s latest customer communication says V4 Pro remains available until September 14, 2026 at 12:00 Beijing time. After that cutoff, requests sent to V4 Pro are expected to route to V4.1 Flash and use V4.1 Flash pricing.
That coexistence window makes this a migration decision rather than a simple benchmark comparison. Teams should verify routing, cache behavior and task quality before the cutoff instead of assuming the two models are interchangeable.
V4.1 Flash pricing
Pricing became effective on September 10, 2026 at 04:00 UTC. DeepSeek applies two weekday peak windows: 01:00–04:00 UTC and 06:00–10:00 UTC, Monday through Friday. All other times are off-peak.
| Price per 1M tokens | Off-peak | Peak |
|---|---|---|
| Cache-hit input | $0.003 | $0.006 |
| Cache-miss input | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Do not compare only the $0.15 input headline with another provider. A cache-heavy workload can be much cheaper, while a weekday peak workload costs twice the off-peak rate. The DeepSeek API cost calculator on APIpulse now exposes both periods and cache states.
What happens to V4 Pro?
| Date | Operational status |
|---|---|
| Through Sep 14, 2026, 12:00 Beijing time | V4 Pro remains available at its existing price |
| After the cutoff | DeepSeek says V4 Pro requests will route to V4.1 Flash at V4.1 Flash pricing |
This is a vendor-announced routing change. Treat it as planned until the cutoff has passed and your own production requests confirm the behavior. If model identity matters for audit logs or evaluations, record the returned model and do not rely only on the requested alias.
Is V4.1 Flash better than V4 Pro?
DeepSeek says V4.1 Flash surpassed V4 Pro across its tested metrics. That is useful release evidence, but it is not an independently reproduced head-to-head result. Different prompts, tool loops and long-context workloads can produce different outcomes.
Run an evaluation set drawn from your own production traffic. Measure task success, structured-output validity, tool-call recovery, total cost per completed task, latency in both price windows and cache-hit rate. The AI Testing & Evaluation hub explains how to build that release gate.
Migration checklist before September 14
- Add V4.1 Flash as an explicit model in staging.
- Replay representative V4 Pro requests and compare task-level outcomes.
- Test both cache-hit and cache-miss traffic.
- Budget for the peak rate rather than assuming every request lands off-peak.
- Confirm retry and fallback logic does not duplicate expensive tool actions.
- Log requested and returned model identifiers around the cutoff.
- Keep a rollback route to a separately supported model if the automatic routing changes behavior.
Which should you use?
Use V4.1 Flash now when low cost and the post-cutoff route are more important than preserving V4 Pro behavior. Keep V4 Pro during the remaining window only when you need time to finish evaluations or preserve a known production baseline.
The safest default is not to wait for silent rerouting: test V4.1 Flash explicitly, update budgets for its time-dependent rates, and deploy through a measured canary. The DeepSeek V4 API guide covers the wider integration, while the AI Application Architecture hub covers routing and fallbacks.
FAQ
Is DeepSeek V4 Pro already retired?
No. DeepSeek’s latest communication says it remains available until September 14, 2026 at 12:00 Beijing time.
Will V4 Pro requests fail after the cutoff?
DeepSeek says they will route to V4.1 Flash rather than fail, but production teams should verify the returned model and behavior after the transition.
What is the cheapest V4.1 Flash rate?
Off-peak cached input is $0.003 per million tokens. Cache-miss input is $0.15 and output is $0.60 off-peak.