What Your AI Vendor Owes You Before a Price Change

AI vendor price-change notice is a contract term, not a courtesy — and four major API providers commit to very different things. OpenAI’s Services Agreement makes pricing-page changes effective 14 days after posting. Anthropic’s Commercial Terms and Google’s Gemini API terms both work on 30 days. DeepSeek’s pricing page commits to no notice period at all: it “reserves the right to adjust” and advises customers to check the page regularly.

That contrast stopped being theoretical this week. On August 11, Anthropic cancelled a price increase it had scheduled weeks in advance. Two days later, on Thursday, August 13, DeepSeek announced a full repricing of its API — effective this Sunday, August 16, at 16:00 UTC, roughly three days after the announcement. Same market, same week, opposite directions, and wildly different amounts of warning.

This guide covers the notice clauses side by side — assembled from each vendor’s live terms, a comparison we have not seen collected elsewhere on this site or in the vendor docs we fetched — then walks through the DeepSeek repricing as the worked case of what “no notice period” looks like in practice, and closes with the operating posture that follows: multi-provider routing, budget bands instead of point estimates, and price checks on a cadence rather than on trust.

The first move was a reversal. On August 11, Anthropic cancelled the Claude Sonnet 5 step-up it had posted at the model’s June 30 launch: the note now standing on its pricing docs states that the $2/$10 per-million-token introductory rate “is now the standard price” and that the previously scheduled September 1 increase to $3/$15 “will not occur.” We covered that cancellation — and the budgeting rule it validates — in our guide to budgeting agents on introductory pricing, so this post will not re-derive it. What matters here is the shape: a dated, published price schedule that simply did not execute.

The second move went the other way. DeepSeek’s API changelog carries an entry dated August 13 announcing that, alongside the V4 family’s general availability, API pricing will be restructured into peak and off-peak rates — effective August 16 at 16:00 UTC. Sections 03 and 04 take that apart in detail, because the arithmetic is not what the announcement language suggests.

A third schedule is still in flight: Google’s Gemini API pricing page bills the Flash line at $0.75/$3.75 per million tokens on the standard tier through December 31, 2026, with $1.50/$7.50 posted to take effect January 1, 2027. As of writing that step-up is posted, not executed — and after August 11, nobody should treat “posted” and “executed” as the same thing in either direction. (There is also a fourth shape — repricing triggered by request size rather than by calendar, like the 200K and 272K long-context thresholds — which we cover separately.)

Three vendors, three different relationships between the posted schedule and what actually bills. That is the sense in which AI price schedules are soft commitments — and it is exactly why the question that matters at signing time is not “what is the price?” but “what does the contract oblige the vendor to do before the price changes?” That question has a precise, checkable answer for each provider, and the answers diverge more than the rate cards do.

On August 14 we pulled the live terms documents for the four providers compared here, and extracted the clause that governs price changes in each. The table below is the side-by-side. Note what it is comparing: not prices, but the contractual minimum warning each vendor owes its API customers before a new price takes effect.

The exact wording matters, so here it is. OpenAI’s Services Agreement, section 6.6: “Price changes on the Pricing Page will be effective fourteen days after they are posted.” The same section also reserves the right to “correct pricing errors or mistakes even after issuing an invoice or receiving payment” — a separate, sharper power worth knowing about. Google’s Gemini API terms: “Google may make changes to this pricing from time to time, effective 30 days after they are posted unless otherwise specified.” DeepSeek’s pricing page, in full: “Product prices may vary and DeepSeek reserves the right to adjust them. We recommend topping up based on your actual usage and regularly checking this page for the most recent pricing information.”

Two refinements before you file this table. First, OpenAI’s agreement defines fees as governed by an Order Form where one exists, with the Pricing Page as the fallback — meaning enterprise customers with a negotiated Order Form can contract for different (usually stronger) protection, and the 14-day clause is what you get without one. Second, these clauses set a floor, not a practice: Anthropic gave far more than 30 days of forward notice on the Sonnet 5 schedule it later withdrew. The floor is what you can enforce; everything above it is vendor goodwill.

The asymmetry at the bottom of the table is the finding. Among the terms we fetched, three vendors commit to a defined warning window; DeepSeek commits to self-service monitoring. That is not an abstract legal nicety — it priced in, concretely, this very week.

Here is what “no notice period” looked like in practice. In early August, an undated footnote appeared on DeepSeek’s pricing page saying the company planned to raise overall API pricing in the near future and describing the expected increase as significant. No rates, no dates, no effective time — a warning with nothing a budget owner could act on beyond generalized unease.

The concrete version arrived on Thursday, August 13, in a changelog entry announcing the V4 family’s general availability. That footnote is gone from the live pricing page; in its place is an exact schedule: peak and off-peak rates for both API models, peak hours defined as 01:00–04:00 and 06:00–10:00 UTC, off-peak set at half of peak, effective Sunday, August 16, at 16:00 UTC. (Converting the published UTC windows to DeepSeek’s home timezone is simple arithmetic: they correspond to 09:00–12:00 and 14:00–18:00 China Standard Time, UTC+8 — that is, Chinese business hours.)

Track the timeline from a customer’s seat. The actionable information — exact rates and an effective time — spans roughly three days before billing changes, and that window covers a weekend. No contract was breached: DeepSeek’s terms promise exactly this, which is to say nothing. Under OpenAI’s clause the same change would have carried a contractual 14-day runway; under Anthropic’s or Google’s, 30. That difference — between three days of practical warning and thirty days of enforceable warning — is what the notice-period table in section 02 measures, and it is invisible on every rate card.

The mechanics of the new schedule itself — how the windows work, what they mean for scheduling batch workloads into cheap hours, and how GLM’s coding-plan off-peak system compares — get their own full guide. What belongs in this post is the procurement lesson: the arithmetic of what the restructure did to the price level.

The announcement frames the change as resource allocation, and the off-peak tier as the encouraged, cheaper option. Both true — off-peak is half of peak, by construction. But the comparison that matters to anyone already running on DeepSeek is not off-peak versus peak. It is the new schedule versus the flat rates billing today. We computed that comparison from the two rate tables on DeepSeek’s own pricing page: the current flat rates, and the posted schedule that replaces them on August 16.

Read the bars carefully: every one of them is above 1.00x. The off-peak tier — the discounted half of the new schedule, the rate you get by scheduling work into the cheap hours — lands between 1.52x and 6.07x of today’s flat rate depending on the line item. And peak hours bill exactly double the off-peak figures, which puts peak rates at roughly 3x to 12x today’s flat rates across the six line items. There is no hour of the day, after August 16 at 16:00 UTC, at which any DeepSeek API line item costs what it costs today.

This is why the announcement’s framing deserves the attention we gave it in section 03. “Off-peak prices set at half of the peak-hour prices” is accurate, and it describes a discount — relative to peak. Relative to the price customers are paying at the moment of the announcement, the same schedule is a universal increase, with the off-peak floor simply the smaller of two rises. Neither statement contradicts the other. Only one of them appears in the announcement, and only the other one belongs in your budget.

The procurement takeaway generalizes past DeepSeek: when a vendor announces a pricing restructure — new tiers, new windows, new dimensions — the direction of the change is not given by the announcement’s framing. It is given by recomputing your own workload under the new schedule against what you pay now. A restructure can hide a rise, a cut, or both at once in different line items, and the vendor is under no obligation to compute the comparison for you.

If the contract gives you three days of practical warning, the defense is not a better spreadsheet — it is an architecture in which no single vendor’s pricing decision is load-bearing. Multi-provider routing layers already exist, are mature, and were built for a closely related problem: provider outages and rate limits. The same machinery that reroutes around a provider that is down can reroute around a provider that has become expensive.

One honesty note before the specifics: in the documentation we fetched, both vendors below market these features as availability and reliability tooling — neither frames them as a hedge against a provider repricing. That framing is our analysis. It follows directly from the same architecture, but you should know which claims are the vendors’ and which are ours.

The price-hedge reading of these features is straightforward. A router with a provider priority list — the order, only, and sort controls documented in Vercel’s AI Gateway provider options, or the allowed_models patterns and cost_tier preferences in OpenRouter’s model-routing docs — means the answer to “DeepSeek reprices on three days’ notice” is a configuration change, not a migration project. Session stickiness (keeping a multi-turn conversation on one provider while routing freely across task types) removes the last operational excuse for hard-wiring a single vendor. The switching cost you would otherwise pay under time pressure gets paid once, up front, calmly.

We made a version of this argument for availability risk in our single-model-risk continuity checklist — a post about what happens when a model becomes unavailable. The point here is different but rhymes: the same dual-source posture that protects you from a vendor’s outage also decouples your spend from a vendor’s price schedule. One piece of architecture, two distinct risks retired. For vendors at the zero-notice end of the section 02 table, we would treat that architecture as a precondition for production dependence, not an optimization.

The financial-planning counterpart to routing is the budget band. A point estimate — “model spend will be $X per month” — encodes an assumption this week just falsified: that posted prices are fixed inputs. A band prices the softness in explicitly. The floor is your workload at current posted rates. The ceiling is the same workload at every announced schedule that could execute inside the budget period — the Gemini January step-up, the DeepSeek August 16 schedule at your actual peak/off-peak mix. And for any vendor with no notice commitment, the ceiling gets a stress case on top, because the announced schedule is not the outer bound of what can change inside a quarter.

Two supporting practices make the band real rather than decorative. First, put price verification on a calendar cadence — monthly for vendors with 30-day notice clauses, weekly for zero-notice vendors. This is not paranoia; it is literally what DeepSeek’s terms instruct customers to do (“regularly checking this page”), promoted from fine print to operating procedure. Second, keep genuinely deferrable work on the batch lane: batch pricing is both cheaper at rest and a natural shock absorber, since deferrable volume is the easiest spend to pause or re-route in the days after an unfavorable announcement.

Looking forward, we would expect the softness this post documents to become more common, not less. Time-of-day windows, promotional intro rates, context-length thresholds, and cache-tier spreads all give vendors more pricing dimensions than the flat per-million-token era had — and every added dimension is another axis a restructure can move along without the headline number appearing to change. Procurement practice built for flat rate cards will systematically misread that landscape.

Pulling the thread together, here is the stance we would take into any AI vendor evaluation where API spend is material:

• Locate the notice clause before you sign. Not the rate card — the clause. Ask where price-change notice lives in the contract, what the window is, and whether it distinguishes increases from restructures. If the answer is “none stated,” price that in as real risk, not as a footnote.
• Ask whether an Order Form can strengthen it. OpenAI’s 14-day clause, for instance, is the default for customers without a negotiated Order Form. At enterprise volume, notice windows and rate locks are negotiable terms — but only if you raise them at signing.
• Treat zero-notice vendors as variable-rate suppliers. Nothing about a vendor’s contractual silence makes it unusable — DeepSeek’s rates remain low in absolute terms even after August 16. It makes the vendor variable-rate: suitable for routed, capped, deferrable workloads; unsuitable as a hard-wired single source.
• Recompute restructures yourself. When a vendor announces new pricing dimensions, run your own workload under the new schedule against your current bill before accepting any directional framing. The section 04 arithmetic took one spreadsheet and an afternoon.
• Calendar the posted step-ups. Every announced future change — the January 1 Gemini Flash-line step-up chief among them right now — goes on the budget calendar twice: once at announcement, once a few weeks before execution, when you check whether it still stands.

This checklist slots into the broader signing-stage diligence — training-data rights, retention, deprecation policy, exit terms — that our AI procurement question list covers in full; the notice clause is one row of that larger review, but after this week it is the row we would check first. And if your team is standing up model spend governance from scratch — routing layer, budget bands, vendor review cadence — that is exactly the kind of operating structure our AI transformation engagements are built to install.