API Price Sunset Calendar: Three Promotional Rates End Within 41 Days

Between November 21, 2026 and January 1, 2027, three promotional LLM API rates reach their end dates, and two of them double. A fixed workload split across GPT-5.6 Sol, Gemini 4 Argon and Gemini 3.8 Flash costs $13,500 a month at today's rates and $22,000 in January, with no change in usage. That is a 63% rise. Every API price sunset on this calendar is published on the vendor's own pricing page, which means the January invoice is already knowable today.

The short answer. To budget for promotional API pricing that expires, give every rate a start date and an end date. Add a follow-on row at the published successor price, and cost each month by the number of days each row covers. Where the vendor publishes no successor price, assume the pre-promotion list rate and label it as an assumption.

The API price sunset calendar for the next 77 days

The inclusion rule for this calendar is narrow on purpose. A rate qualifies only if the vendor's own pricing page carries a date on it, and only if I opened that page on October 16, 2026. Aggregator listings that restate a date did not count. Neither did a launch blog post whose wording the pricing page has since replaced.

Under that rule, the five largest text-model vendors carry three live dated rates, one recurring daily rate, and one introductory rate that was cancelled before it bit. The full Gemini Flash schedule, including the short life of 3.7 Flash, is covered in the breakdown of the Gemini Flash January price step. This post adds the Argon cliff and prices all of them together.

The GPT-5.6 Sol "launch price" is from the Digital Applied Q3 tracker, which records the model at $5 and $30 on July 9, 2026. OpenAI's page describes the promotion as a 20% input and 33% output reduction, which resolves to the same figures. xAI's model page and Anthropic's current price table carry no dated rate of any kind.

Gemini 4 Argon: the cliff most budgets have not seen yet

Argon is the entry most teams will miss, because the date is recent. Third-party pricing write-ups from early October said Google had not disclosed how long the introductory period would last. The pricing page, last updated October 15, 2026, now answers it in every row.

Standard input goes from $2 to $4 per million tokens. Output, which includes thinking tokens, goes from $10 to $20. Priority goes from $3.60 and $18 to $7.20 and $36. Caching is implicit only on Argon, and the cache rate doubles from $0.10 to $0.20 alongside everything else. Nothing on the page escapes the date.

GPT-5.6 Sol: a floor, not an expiry

OpenAI's wording is different in kind. The GPT-5.6 Sol model page says the promotional pricing "is available at least through November 21, 2026". That commits OpenAI to nothing after the date, in either direction. The page publishes no successor rate.

I would budget Sol at $5 and $30 from November 22 and treat any extension as found money. The reason is simple. An extension costs you nothing if you planned for the rise, and a rise costs you a month of overrun if you planned for an extension.

What 28 API price changes in one quarter tell you

The Digital Applied tracker counted 28 dated changes to standard-tier text-model API prices between July 1 and September 30, 2026. They span Anthropic, OpenAI, Google, DeepSeek and xAI. It counts new models at their launch price, so the number measures pricing activity, not 28 repricings of models you already use.

The 28 is the wrong number to worry about. A buyer who reads it as "prices are volatile" will conclude the answer is a bigger contingency line. The more useful reading is that most changes are announced with dates, and dated changes can be modelled exactly.

Most of the movement was downward

New flagships mostly launched at or below the model they replaced. GPT-6.1 Sol launched on September 29 at $2 and $10, half the promotional rate on GPT-5.6 Sol. Anthropic's API release notes record Opus 5.5 at $4 and $20 against $5 and $25 for Opus 5. They also record an October 7 cut to Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens.

That direction matters for budgeting. The cliffs on this calendar sit on older or promotional models, while the successor models are often cheaper at list. The cheapest response to a cliff is frequently a migration, which is a cost of its own, covered further down.

The one vendor that raised, and how

DeepSeek was the only vendor in the tracker whose prices went up in the quarter. On August 16 it replaced a flat rate with peak and off-peak billing, and V4 Flash moved from $0.14 and $0.28 to $0.44 and $1.32 at peak. V4.1 Flash then replaced it on September 10 at $0.30 and $1.20 peak. A price rise can arrive as a change in billing structure rather than a new number, which is the subject of the analysis of why cheaper tokens still produce rising AI bills.

The API price cliff on a fixed workload

The calendar tells you dates. A budget needs a figure. So here is a deliberately plain workload, priced on list rates with no caching, standard tier and short context.

Each model runs 1 billion input tokens and 200 million output tokens a month. That 5 to 1 input to output ratio is a common shape for retrieval-heavy assistants. Volume is held flat across all four months so the only thing moving is price.

Author's calculation from vendor list prices read 16 October 2026. November assumes the Sol promotion ends on November 21, so 21 days bill at $4 / $20 and 9 days at an assumed $5 / $30. Every Sol figure after November 21 is an assumption, because OpenAI publishes no successor rate.

Read the October column differently. Of the $13,500, $8,500 a month is price you are borrowing against a date already printed on the vendor's page. That is the gap between October and January, and it is larger than the whole October cost of the Sol workload.

A team that sets its 2027 AI budget from its October invoice will understate the run rate by 39%. That is $8,500 divided by $22,000. The understatement is not an estimation error. It is the promotion working exactly as designed.

Why output-heavy workloads fall harder on the Sol cliff

The Gemini cliffs double every rate, so your input to output mix does not change their size. The Sol cliff is different, because the promotion cut output by 33% and input by 20%. Unwinding it hits output harder.

The arithmetic behind the bars is short. An input-heavy job of 1 billion input and 20 million output tokens goes from $4,400 to $5,600. A 1 to 1 job of 200 million each way goes from $4,800 to $7,000. Run your own ratio from last month's usage export before you pick a contingency figure.

DeepSeek's API price sunset happens every weekday

DeepSeek has no promotion with an end date. It has something more awkward for a budget: a rate that expires and returns twice a day. The DeepSeek pricing page states that off-peak rates are half of peak, and that peak runs 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays.

For the same 1 billion input and 200 million output tokens, all-peak traffic costs $540 a month and all off-peak costs $270. Spread evenly across the week, the blended figure is about $326. A team whose users sit in India works through most of the peak window, so its real figure sits much closer to $540 than the even spread suggests.

The budget move is to measure, not assume. Export a week of request logs, bucket tokens by UTC hour, and compute your own peak share. Batch jobs that can wait belong after 10:00 UTC, and that is a scheduling decision rather than a vendor negotiation.

A spreadsheet method for forward-pricing LLM API rates

Most AI budgeting sheets have one row per model and one price per row. That shape cannot represent a sunset, because the price is a property of a date range, not of a model. The fix is a second sheet, and it takes about an hour to build.

Step 1: one row per price period, never per model. GPT-5.6 Sol becomes two rows. The first runs August 21 to November 21 at $4 and $20. The second runs from November 22 at $5 and $30, with the Kind column set to "assumed". Argon and Gemini Flash each get two rows with published successor rates. Sonnet 5.5 and GPT-6.1 Sol get one open-ended row each.

Step 2: let the overlap formula split the month. The days formula in column G does the only clever thing in the sheet. It counts how many days of a given month fall inside each rate's window. A month column header holding the first and last date of the month is enough to drive it, and a SUMIFS on model and month rolls the rows back up into the budget view.

This is the same calculation behind November's $8,900 in the table above. Twenty-one days at the promotional rate plus nine days at the assumed rate, each weighted by its share of a 30-day month. Purpose-built tools do this too, and the comparison of AI spend forecasting tools covers which ones model dated rates natively.

Step 3: add a cliff line to the budget view. Below the monthly total, add one row: the difference between what the month costs and what it would cost at each model's latest dated rate. In October that line reads $8,500. It is the size of the price you are borrowing, and it should shrink to zero on January 1.

Presenting it this way changes the conversation with finance. A CFO who sees a separate line marked "promotional, expires on these dates" approves the higher January number in October. A CFO who sees a 52% jump in December with no warning asks a different question.

Where this budgeting method is weakest

The method prices the dates exactly. Everything else in it is an assumption, and three of those assumptions are shakier than the dates.

Volume is the bigger variable, and it is held flat here

The worked example freezes token volume to isolate price. Real workloads do not hold still. Newer models can spend more thinking tokens per task, and output on Argon and Gemini Flash is billed including thinking. A model swap that saves 50% on the rate can give some of it back in volume, and no rate sheet captures that. Measure tokens per completed task, which is the subject of the guide to building an LLM eval suite.

The Sol successor rate is an inference, and vendors do reverse

The $5 and $30 figure for GPT-5.6 Sol after November 21 is the launch price, not a published successor. OpenAI could extend the promotion, cut the list price, or retire the model. Anthropic showed in August that a scheduled increase can be cancelled outright, when Sonnet 5 stayed at $2 and $10. Budgeting at the higher figure is the conservative choice, not the predicted one.

List price is not what every buyer pays

Everything above uses public list rates. Enterprise agreements, committed-use discounts, cloud marketplace pricing and regional uplifts all sit outside it, and the tracker behind the 28-change count excludes them too. If your contract fixes a rate, the vendor's calendar may not apply to you at all. Check the contract before you book the cliff. The 28 figure itself is a third-party count and should be read as directional, although every dated row I used in this post was checked against the vendor's own page.

Migrate, absorb or lock: the decision at each cliff

Each sunset leaves three options, and the right one differs by cliff.

Migrate when a cheaper successor exists. GPT-6.1 Sol lists at $2 and $10, so the Sol workload above would cost $4,000 a month on it against $11,000 on post-promotion Sol. Migrating that one workload cuts the January bill from $22,000 to $15,000. That saving pays for a real evaluation run, and skipping the evaluation is how teams trade a pricing problem for a quality incident. The switching mechanics are in the piece on model switching costs.

Absorb when the post-cliff rate is simply the market rate. Argon's December price of $4 and $20 matches the list price of Claude Opus 5.5 on Anthropic's pricing page and the current promotional price of GPT-5.6 Sol. In my view the Argon introductory rate was a launch discount, not a new price level, and the $4 and $20 figure is the one to plan on.

Lock when your volume is large enough to negotiate. A committed-use agreement that writes the current rate in as a ceiling for 12 months converts a vendor's calendar into your contract. The renewal dynamics that follow such deals are covered in the analysis of grandfathered pricing at renewal. I would not lock a Flash-tier workload, because the absolute dollars are small and successor models keep arriving below the old price.

Frequently asked questions

Where to start before November 21

Pull last month's usage export and find every model string that appears on the calendar above. For each one, record the input to output ratio, because it decides how big the Sol cliff is for you and nothing else does.

Then build the rate segments sheet with two rows for each of those models and read the December total. If that number surprises you, you have five weeks to run an evaluation on the successor model, which is more time than most teams give a migration and less than it feels like.

Related reading