LLM price floors vs promotional cuts

A token budget needs a rate that survives the next invoice, not a headline that expires. The trade-off is direct: a promo can lower a line, but it can also end or skip your plan. On August 16, 2026, DeepSeek V4 Pro moved from a flat $0.435 input and $0.87 output per-million-token rate to a split peak/off-peak schedule. DeepSeek's off-peak output price became $1.98 per million tokens, above the earlier $0.87 flat output price. The budget line should carry the rate, the window, the plan, and any expiry.

On August 21, 2026, OpenAI cut GPT-5.6 Sol's list price from $5 input and $30 output to $4 input and $20 output, a 29% blended-price drop. That cut was promotional and was stated to be available at least through November 21, 2026.

The equal-weight frontier index fell 5.7% in August 2026 and was 9.4% below its February 23, 2026 first reading. The index shows direction, not invoice stability. The question is which rate survives a plan change, a schedule change, and a promo end date.

The floor test has three lines:

• Mark prices with expiry or promotional wording as temporary.
• Model peak/off-peak and plan coverage instead of the headline rate.
• Treat only non-promotional, plan-wide, schedule-stable prices as floors.

Classify each move in order. An expiry or promotional wording makes it a temporary promo. A peak/off-peak change makes it a schedule change, so rebuild the blended rate. A non-promotional, plan-wide, schedule-stable price is a durable floor. That keeps the budget from absorbing a promo as a floor or a schedule change as a cut.

Promo wording is a temporary assumption

A price that depends on promotional wording that can be revised, such as OpenAI's GPT-5.6 Sol price of $4 input and $20 output after November 21, 2026, is a promotional price rather than a permanent floor. Keep the promo in a separate column, with the end date next to the rate; keep the stable list price in the floor column.

Schedules change the arithmetic

Model peak/off-peak windows before the headline rate. If traffic lands off-peak, the lower headline can raise the blended rate. A schedule change is not a price cut until the traffic-weighted rate is lower.

Averaging the two rates without traffic weights is the mistake. The arithmetic is the traffic split: blended output rate = peak share × peak rate + off-peak share × off-peak rate. The finished model has two rates, a traffic split, and a time window for each rate.

Coverage decides applicability

OpenAI's GPT-5.6 Sol discount applied to pay-as-you-go usage, Codex credits, and qualifying Work plans, but not to Pro, Plus, or Business subscriptions. If production spend sits on an excluded subscription, the headline cut is not your cut. A coverage-limited cut is not a plan-wide floor until every production plan is included.

Weight the cut by the share of tokens on covered plans. The finished table has three columns: stable rate, promo rate, and coverage.

Invoices reveal the floor

Reconcile the floor against the invoice before the next budget review. Pull line items by model, plan, and rate window. Compare the stable rate to the billed rate, then compare the promo rate to the billed rate. If the promo ended, the stable rate should reappear without a manual edit. Treat a mismatch as a data problem, then fix it.

The artifact is a variance note, not a guess. If the invoice shows a lower rate after the promo window, flag it as a new promo or a billing error. A bill at the old list price means the floor held. A rate that matches neither means the model needs a new input.