Cut inference bills 77-80%: a migration checklist

If you run DeepSeek V4-Pro in production, the provider will reroute all V4-Pro API calls to V4.1-Flash on September 14, 2026, and bill them at V4.1-Flash rates until a V4.1-Pro model is available. The trade-off is lower token cost against peak-hour exposure, KV memory, and concurrency: V4.1-Flash is 77% to 80% cheaper than its predecessor and raises concurrency from 500 to 2,500 requests. The arithmetic starts with off-peak cache-hit input at $0.003 per token, versus $0.022 for outgoing V4-Pro. Peak-hour prices are twice off-peak prices.

The price spread is the entry fee

Peak and off-peak billing changes the average. The schedule, effective September 10, 2026, bills weekdays 01:00–04:00 and 06:00–10:00 UTC as peak; everything else is off-peak. If your load is mostly off-peak, the effective rate sits near the low end; if it is concentrated in those weekday windows, the discount shrinks.

Compute the effective rate before you trust the headline. Weight cache-hit and cache-miss input by the share of input tokens that hit the cache, then apply the peak/off-peak mix to the result. Output tokens often dominate the bill for long-running agents, so a small output-rate change can move the total more than a large cache-hit change. Do not use a single blended rate. Your cache-hit share depends on prompt reuse, system-prompt length, and session restarts. Your peak/off-peak share depends on user geography and batch jobs. If either mix shifts, the effective cost per token shifts with it.

KV cache and concurrency decide capacity

Cache behavior decides how much of the input bill you actually pay. V4.1-Flash uses an 890-byte-per-token KV cache, about 3.9 times smaller than the previous generation and more than 400 times smaller than DeepSeek-V1. Smaller KV state means more concurrent sequences fit in the same GPU memory.

The model has 552 billion total parameters, with 8 billion active for input and 16 billion active for output. That split keeps per-token compute low while leaving room for long contexts. If your old stack was limited by KV memory, the smaller cache changes how many sessions you can keep alive before adding GPUs.

Measure memory per sequence at your longest context, then compare the resulting concurrency headroom against your peak request rate. A model that is cheaper per token but forces you to buy more capacity can still lose on cost per completed task. Also check batch versus interactive traffic. Batch jobs can be shifted into off-peak windows, while interactive traffic may be pinned to peak hours. The same model can be a win for batch traffic and a loss for interactive traffic.

Benchmarks set the quality gate

A cheaper model is only a win if your tasks still pass. In max-reasoning tests, V4.1-Flash scored 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1, and 88.1 on CyberGym, ahead of V4-Pro. DeepSeek said V4.1-Flash outperformed V4-Pro in capability, price, speed, and completion time in internal and external tests.

Your gate should be narrower: your eval set, your latency SLO, and your failure budget. Run the same prompts, tools, and failure cases you use in production, then compare pass rate, latency, and cost per completed task. A benchmark win does not clear your SLO if your traffic is tool-heavy, long-context, or latency-bound. The quality gate should stop the migration before users see the regression.

Route traffic only with a rollback trigger

The routing change makes the migration less optional for existing traffic. Use the checklist below before you flip traffic. Each item maps to a cost or failure mode that can erase the savings.

• Compute effective cost per token from cache-hit/miss and peak/off-peak rates.
• Compare tokens/sec and GPU-hour against your current baseline.
• Measure KV cache memory per sequence at your max context length.
• Verify concurrency headroom against your peak request rate.
• Run a quality gate on your production eval set.
• Set a rollback trigger on latency, error rate, or eval regression.

Keep the old endpoint available until the new model clears the gate. If the effective cost per completed task rises, or if latency and error rates cross your thresholds, roll back before the next billing cycle.