In August 2026, three labs turned an LLM’s price into a moving target
In two weeks of August 2026, DeepSeek introduced peak/off-peak billing, Google launched a tier whose price doubles in January 2027, and Anthropic cancelled a planned increase. For anyone budgeting inference spend, the per-token price is no longer a fixed number but a three-variable equation.
August 13, 2026. DeepSeek announces a peak/off-peak rate card for its V4 API, effective August 16. August 13, 2026. Google launches Gemini 3.7 Flash at an introductory rate that doubles on January 1, 2027. August 11, 2026. Anthropic cancels the planned increase for Claude Sonnet 5.
Three labs, three decisions, one direction: in August 2026, the published price of an LLM stopped being a single number. For a CTO or an SRE budgeting inference spend, the consequence is immediate — a quote built on today’s rate card can be wrong within five months.
DeepSeek, or the price that depends on the hour
DeepSeek chose the most radical approach. Since August 16, 2026, its V4 API bills differently by the hour: a peak rate during the 01:00–04:00 and 06:00–10:00 UTC windows, and an off-peak rate everywhere else, set at half the peak rate.
Presented as a discount, the change is actually a broad price increase. V4 Pro currently costs $0.435 input and $0.87 output per million tokens. After the switch, the off-peak rate rises to $0.66 and $1.98, and the peak rate to $1.32 and $3.96. On output, the cheapest hour of the new schedule costs 2.3× today’s price, and the most expensive 4.6×.
The number that matters is therefore not “off-peak at −50%” but “even the cheapest hour costs more than double today’s rate”. A workload that runs around the clock — a nightly batch, a background agent, a classification pipeline — will see its DeepSeek bill climb regardless of the hour.
Google, or the price with an expiry date
Google went the other way, but with a clock attached. Gemini 3.7 Flash launched on August 13, 2026 at $0.75 input and $3.75 output per million tokens. The pricing page states plainly that this is an introductory rate: on January 1, 2027, it becomes $1.50 and $7.50 — a twofold increase.
There is an immediate precedent. Gemini 3.6 Flash, released in late July, was cut to that same $0.75/$3.75 about three weeks after launch — a 50% drop that also expires at the end of 2026. Two models, two generations, one mechanism: an aggressive price to fill the pipeline, then a scheduled climb back up on a fixed date.
For a buyer, the lesson is simple. The displayed Gemini Flash rate is a temporary rate. A contract, budget, or migration built on $0.75/$3.75 must include the scenario where that price is double in four months. The expiry date is part of the price.
Anthropic, or the increase that never happened
Anthropic did the rarest thing of the three: it cancelled a price increase before it happened. Claude Sonnet 5, launched on June 30, 2026 at $2 input and $10 output, was sold as an introductory rate through August 31, reverting to $3/$15 on September 1.
The pricing page now says the opposite: the $2/$10 “is now the standard price” and “the previously scheduled increase to $3/$15 on September 1, 2026 will not occur”. Any team that planned a migration off Sonnet 5 to avoid a 50% rise in September can stop.
The decision is best explained by competition. Facing a Gemini Flash at $0.75 and an ecosystem of open models whose prices keep collapsing, holding a premium rate on the “workhorse” model would have been a risky bet. Anthropic preferred to lock in volume on Sonnet 5 rather than defend a theoretical margin.
A rate card is no longer a rate card
What the three decisions draw together matters more than any of them alone. For years, an LLM’s price fit in two columns: input and output, per million tokens. In August 2026, that card decomposed into three variables, which the three labs each explored separately.
DeepSeek introduced the temporal dimension — the hour of the call becomes a cost factor. Google introduced the calendar dimension — the advertised price is only valid until a date. Anthropic introduced the reversibility dimension — an announced increase is no longer a certainty. None of the three has combined all three yet, but the direction is clear.
The consequence for teams is concrete. Inference cost control must now model scenarios rather than memorize rates. A budget built on August’s card must ask: what happens if the rate doubles in January, if processing lands in peak hours, or if the advertised discount is later withdrawn?
Verdict
If your inference spend is cost-sensitive, model your bill across several horizons rather than a single rate. For DeepSeek V4, assume the real price sits somewhere between off-peak and peak, and redo your overnight batches — off-peak will not shield them from the increase. For Gemini Flash, treat $0.75/$3.75 as an introductory rate and budget the January 2027 doubling. For Claude Sonnet 5, stay put: the announced increase will not happen, and $2/$10 is now the reference price.
If you are negotiating an annual contract, demand that the rate card be guaranteed over time. August’s moves show a published price can change meaning within two weeks — by the hour, by the date, or by cancellation. A rate-freeze clause is now worth more than a one-off discount.
The whole picture fits in one sentence: an LLM’s price is no longer a point but a trajectory — and a budget that models only the point is already wrong.
References
- Capital & Compute — New AI models released in August 2026: the full list, verified August 25, 2026
- DeepSeek — API pricing, accessed August 25, 2026
- Google — Gemini API pricing, accessed August 25, 2026
- Anthropic — Pricing, accessed August 25, 2026
- Google — Introducing Gemini 3.7 Flash, August 13, 2026