Re-run your rate card: endpoint, cache, and window decide cost

You can keep the model pin stable, but the bill moves when the serving endpoint, cache state, or peak window changes. An agent loop sends cached input, uncached input, and output tokens. Multiply each slice by its rate, then add the products. The model name routes the request; the rate card prices it. Do this before you trust the bill.

Endpoint, cache state, and window set the price

DeepSeek's new V4.1 Flash pricing schedule took effect at 04:00 UTC on September 10. The off-peak cached input price is $0.003 per million cached input tokens, and the off-peak output price is $0.60 per million output tokens. Peak rates are double those off-peak rates.

Beginning at 04:00 UTC on September 14, DeepSeek will route requests sent to V4-Pro to V4.1-Flash and bill them at Flash rates until V4.1-Pro arrives. DeepSeek said the V4-Pro endpoint would stay up, with its requests processed by V4.1 Flash and billed at Flash rates.

The audit makes the bill testable

Run this audit when a provider changes pricing, routing, or model availability. Each step is small enough to finish before the next agent loop starts.

• Resolve the pinned model ID to the endpoint that actually serves it.
• Split the last billing period into cached input, uncached input, and output tokens.
• Tag each token slice by peak or off-peak window.
• Apply the current rate card to each slice, not to the model name.
• Compute the bill delta for an off-peak shift on a safe slice.

Start with the endpoint check: a pin can keep its old name while the serving model changes. The token split exposes the price difference between cache states. Peak and off-peak carry different multipliers, so the window tag is mandatory. The rate card applies to slices, since the invoice is a sum. A forecast is only as good as the bill delta that tests it.

The old-rate spread ranks the workloads

Against V4 Flash off-peak rates of $0.007 per million cached input tokens, $0.22 per million uncached input tokens, and $0.66 per million output tokens, V4.1 Flash cuts those lines by about 57%, 32%, and 9%. For workloads moving from V4 Pro off-peak rates of $0.022 per million cached input tokens, $0.66 per million uncached input tokens, and $1.98 per million output tokens, the corresponding prices fall by roughly 86%, 77%, and 70%. Rank workloads by that spread before you move them. Start with the workload that has the largest drop.

Peak and cache state decide the shift

The window changes the multiplier on the same token slice. If an agent loop runs in batches, the window can matter more than the model label. Record the start and end time of each request, then group usage by window before applying rates.

For long-running agents, cache state is the missing variable. A prompt can be cached in a call and uncached in the next, and the slices price differently. Keep the cache-hit rate in the token budget, not just the total token count, because the cached input slice can dominate the token count.

Pin the endpoint you can verify, not the model name you prefer. If a provider keeps an old endpoint alive during a transition, your config may still pass while the serving model changes. The audit should record the actual serving model for each pin, the cache state for each token slice, and the window for each request.

Test the off-peak shift with a bill delta, not a dashboard estimate. Move a safe slice of traffic, hold the token mix constant, and compare the invoice before and after. If the delta does not match the rate-card arithmetic, the cache state or window tag is wrong. The delta is the bill drop estimate.