The Same Prompt, Four Different Bills: DeepSeek Now Charges by the Clock, Grok by Length, OpenAI by the Calendar (August 2026)

In nine August days, the list price quietly stopped being a number. On August 13 DeepSeek shipped V4-Pro GA and, three days later, switched its API to peak and off-peak billing: the same output token costs $3.96 per million at 09:00 UTC on a Tuesday and$1.98 on a Sunday. On August 12, xAI launched Grok 4.6 with a threshold that doubles the price of the entire request the moment your prompt crosses 200K tokens. And on August 21, OpenAI cut GPT-5.6 Sol by 20% or more, but labeled the new price promotional, valid through at least November 21. Three vendors, one message: what a prompt costs now depends on when you run it, how long it is, and what week it is. The price moved from the model's page to your prompt's metadata, and if your prompts live in scattered strings, you cannot see it moving.

What changed, in nine days

The facts first, because each one looks small until you stack them.

DeepSeek released V4-Pro GA on August 13: 1M context, output up to 384K tokens, three thinking-effort levels (low, high, max), native Responses API support. Then, effective August 16, it introduced time-of-day billing. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday; everything else, weekends included, is off-peak at exactly half price. V4-Pro output runs $3.96 per million tokens at peak and$1.98 off-peak; cache-miss input is $1.32 versus$0.66. When we covered the July retirement of the old deepseek-chat and deepseek-reasoner names, V4-Pro listed a flat $0.435 input and$0.87 output. Read those numbers again: the vendor that built its brand on being always-cheap raised its frontier list price roughly 4.5x on peak output, and the discount lever it offers back is a clock.

xAI shipped Grok 4.6 on August 12: 500K context, $2 input and$6 output per million tokens, cached input at $0.50. Until your prompt reaches 200K tokens. From that point xAI bills all tokens in the request at the long-context rate:$4 input, $12 output,$1 cached. Not the tokens past the threshold, the whole request. One retrieval chunk too many and the entire call reprices at 2x, with no error and no header shouting about it.

OpenAI cut GPT-5.6 Sol on August 21 from $5/$30 to $4 input and$20 output per million tokens, across the API, Codex credits, and eligible ChatGPT Work plans; subscriptions stay where they were. The stated window: through at least November 21, 2026. Sol keeps its own length cliff too, inherited from the July launch: past 272K input tokens the request bills at 2x on input and 1.5x on output. A price cut with an expiry date means your unit-economics spreadsheet now has a review date in it.

None of this is exotic anymore. This is the pricing structure of three of the most used APIs on the market, as of this week.

Axis 1: the clock, and whose night it is

Time-of-day pricing is the genuinely new one, and it comes with a trap for anyone west of Shanghai. DeepSeek's peak windows, 01:00-04:00 and 06:00-10:00 UTC on weekdays, map to daytime in China. They also map to the European night. The oldest cost instinct in computing, run the heavy batch at night, is now exactly wrong on this API if your night is Europe's: a cron firing at 02:00 UTC lands in the most expensive window of the week. Off-peak is not your night, it is the vendor's schedule, and the cheapest hours for a European or American team are their own working afternoon and the entire weekend.

That makes scheduling a pricing decision. Anything that is not latency-sensitive, evaluation runs, batch scoring, dataset generation, regression suites, nightly summaries, now has a correct and an incorrect time to exist, 2x apart. The interesting operational question is not "can we save by moving it", it is "which of our prompts are allowed to move". That answer, latency tolerance, is a property of the individual prompt, not of the account.

Axis 2: length, and the cliff you cannot see from the prompt

Grok 4.6 and GPT-5.6 price length the same way: not as a slope but as a cliff, and the fall reprices everything behind you. Grok doubles the full request at 200K tokens; Sol multiplies input 2x and output 1.5x at 272K. We argued when 1M windows first shipped that stuffing context degrades answers long before it hits the limit. The 2026 update is blunter: now it also has a price tag with a step function in it.

The dangerous part is who controls the length. Usually not the prompt author: it is the retrieval layer adding one more chunk, the agent harness dragging in one more file, the conversation history growing one more turn. The prompt that cost $2 per million input yesterday costs$4 today because a loop upstream got more generous, and nothing failed. If a prompt is designed to run near a cliff, the threshold belongs next to the prompt as an explicit context budget, something a reviewer can see and a harness can enforce, not a surprise in next month's invoice.

Axis 3: the calendar, and prices as moves in someone else's game

The OpenAI cut is the least structural of the three, and the most revealing. Reuters attributes it to competitive pressure from Anthropic and the Chinese labs, and the promotional label makes it explicit: this price is a move, not a fact. It can move back on November 21. DeepSeek's hike travels the other direction on the same lesson. We wrote in June that the model is becoming the commodity and the prompt library the durable asset; a quarter later, vendors themselves treat list prices like airline fares. Budgeting a workload on today's number with no review date is planning on weather.

And underneath all three axes sits the one we have already covered twice: effort. Opus 5 hands you a per-request dial that quietly multiplies output cost, GPT-5.6 routes across three tiers with two reasoning dials on top, and V4-Pro now takes low, high or max. Clock, length, calendar, effort: four axes, none of which appear anywhere in the prompt text itself.

Four axes that reprice the same prompt without touching a character of it. The text is identical; the bill is not.

Route by cost: the matrix we actually use

Here is how we route after this month, one row per job shape. The thresholds are ours; steal them and adjust.

Five job shapes, five routes. Every cell is a property of the prompt, not of the vendor account.

What we found in our own library

We ran the same pass on our prompts that we recommended after the DeepSeek deprecation, this time reading for cost instead of model names. Two findings made the pass worth it. The scheduled evaluation job that nearly shipped a dead model name in July turned out to fire at 06:30 UTC on weekdays, which was a perfectly reasonable time when the price was flat and is squarely inside DeepSeek's peak window now; moving the cron to Saturday was a one-line change that halved that job's rate card. And two retrieval-heavy prompts were budgeted by feel, sitting close enough to Grok's 200K cliff that a fatter-than-usual retrieval day would have doubled their cost silently; they now carry an explicit budget and a truncation rule.

The lesson is the one this blog keeps relearning in different clothes: nothing was broken, nothing errored, and both problems were invisible in the prompt text. A prompt string tells you what the model will read. It no longer tells you what the call will cost. That information, target model, effort level, context budget, run window, has to live somewhere reviewable, or it lives nowhere.

The library is where the price becomes visible

This is, bluntly, why we keep prompts in a library instead of scattered across repos and dashboards. A prompt entry that records its target model and tier, its effort level, its context budget and its run window turns each of August's changes into a diff you can grep for: DeepSeek reprices, filter by vendor and check the schedule column; a cliff moves, sort by context budget; the Sol promo expires in November, the review date is already written down. And scoring each prompt closes the loop on the effort axis: score the output at low, then escalate only on a miss, and you find the cheapest gear that passes instead of paying max everywhere out of caution.

The prompt did not change this month. The four prices attached to it did, twice before breakfast in DeepSeek's case. Version the metadata with the prompt, and the next repricing is a search, not a surprise.