On August 16, 2026, DeepSeek switched its API to peak and off-peak billing, charging by the clock like a power utility. On August 23, seven days later, it reversed course: Bloomberg reported that from that morning every weekend hour is billed at the off-peak rate, taking the discounted share of the week to roughly 79% [1]. We know the exact distance between those two dates painfully well, because the first one was example A in the article we published about multi-vendor pricing schemes, and it shipped on the same day the scheme it described stopped being true. That is the whole lesson of this follow-up, and it is a better one than we intended to teach: the price you record next to a prompt is not a property of the prompt. It is a snapshot with an expiry date, and in August 2026 that expiry date can be a week.
What actually changed, in one paragraph
DeepSeek's clock-based scheme had made weekend and night hours cheap and business-hours peak expensive, with the peak rate on deepseek-v4-pro reaching 27 yuan per million output tokens [1]. The August 23 change did not abolish the clock: weekday peak pricing still exists. It declared the entire weekend off-peak, Saturday and Sunday, from midnight Beijing time. If you had rebuilt your batch jobs around the original calendar (as the scheme invited you to do), the arbitrage you engineered got wider; if you had priced weekend work at peak rates in a client quote or an internal budget, your numbers were stale within a week of writing them.
Timeline of seven days: August 16, DeepSeek launches peak and off-peak API pricing; August 23, the weekend becomes entirely off-peak, the same day our multi-vendor pricing article shipped
Seven days between a pricing scheme and its first reversal. The mistake is not writing about a scheme that changes. The mistake is recording a price anywhere, in an article or in a prompt library, without a date on it.
The opposite move, same fortnight
While DeepSeek was loosening its clock, Anthropic did the reverse of a reversal. Claude Sonnet 5 launched in June with introductory pricing of 2permillioninputtokensand10 per million output, scheduled to rise to 3/15 on September 1. On August 11 Anthropic announced the increase would not happen: the introductory price is now the standard price, permanently [2]. The stated reason is itself a lesson in how tangled model economics have become: the newer tokenizer produces roughly 30% more tokens for the same text, so the vendor's original "cost-neutral" arithmetic no longer held, and the sticker price stayed down partly because the effective price per word had quietly gone up.
Want to know how effective your prompts are? Prompt Score analyzes them on 6 criteria.
So in a single fortnight, one vendor rewrote a week-old scheme in the customer's favor, and another canceled a scheduled increase that thousands of teams had presumably budgeted for. Two moves in opposite directions, and both invalidate the same thing: whatever you wrote down about prices in July.
Two vendors, two opposite pricing moves in one fortnight: DeepSeek widens the off-peak window in the customer's favor after 7 days; Anthropic cancels a scheduled increase and makes launch pricing permanent
Neither move was predictable from the announcement that preceded it. A price schedule is a vendor's current intention, not a commitment, and your library should treat it that way.
What this does to the routing playbook
Our pricing article argued that price has become a property you route on: the same prompt costs different amounts depending on the clock (DeepSeek), the context length you feed it (Grok), and the calendar of promotional windows (OpenAI). Nothing in this week changes that thesis. What changes is the maintenance contract that comes with it. If price is metadata you route on, then stale price metadata produces confidently wrong routing, the same way a stale policy inside a support prompt produces confidently wrong replies. Three fields turn a price note from a liability into an asset:
The date you recorded it. "DeepSeek off-peak, 50% discount" is a trap. "DeepSeek off-peak per the Aug 23 schedule, checked Sept 2" is information. The date is what tells the next reader, usually you in October, whether to trust it or re-check it.
The source, as a link. Not a screenshot, not a memory: the vendor's pricing page URL, so re-checking costs thirty seconds instead of a search session.
A review cadence. Prices in this market now change faster than models do. We treat any price note older than a month as expired by default, and this month proved even that may be generous.
The place those fields live matters less than that they exist: a notes field next to the prompt, a metadata block, a comment in the config. What matters is that they travel with the prompt, because a spreadsheet of prices divorced from the prompts that depend on them is exactly the spreadsheet failure mode that loses to any real library: no history, no diff, no way to know which prompt was tuned against which price regime. When the price line changes, we version it like we version the prompt text, and choosing the model for a task starts from metadata that says when it was last true.
The techniques you're reading about work. Test your prompts now with Prompt Score and see your score in real time.
Abstract advice about metadata is easy to nod at and never do, so here is the actual shape of the thing. This is the notes block we keep next to a summarization prompt that routes across vendors; yours will differ in the numbers, not in the structure:
ROUTING NOTES · checked 2026-09-02
- Default: DeepSeek v4 · off-peak covers all weekend + weeknights
per the Aug 23 schedule (source: api-docs.deepseek.com/pricing).
Batch jobs pinned to Sat-Sun. Peak still applies weekday daytime.
- Escalation: Claude Sonnet 5 · $2/$10 now PERMANENT (Aug 11 note:
Sept 1 increase canceled; tokenizer uses ~30% more tokens, so
compare per-task cost, not per-token price).
- Long-context jobs: check Grok's 200K step BEFORE batching
(price doubles past the threshold · see the four-bills piece).
- Review by: 2026-10-02, or on any vendor pricing announcement.
Four lines of routing logic, and every one of them carries the three fields: what was true, where it says so, and when it was last checked. Notice what the block does not contain: per-token prices copied by hand. A copied number is stale the moment the page changes; a dated claim with a link tells you exactly how much to trust it and where to look. The "review by" line is the piece most libraries miss, and it is the one that would have saved anyone who quoted weekend work at peak rates this month.
Two of the four axes from the pricing piece deserve a special flag in your notes because they fail silently. Clock-based pricing (DeepSeek) fails loudly: the invoice line makes it obvious. But a tokenizer change (Anthropic's ~30% denser tokenization behind the frozen sticker price) and a context-length step (Grok's threshold) both change your cost without any price changing, which means no announcement will warn you. For those, the only detection is the per-task cost check in your own usage data, which is one more reason the price note belongs next to the prompt that generates the spend.
The five-minute monthly price audit
The routine that falls out of this is small enough to actually happen. Once a month, or when a vendor makes news: open the pricing page of each provider you route to, compare against the price line recorded next to your routing prompts, and update the ones that moved, with today's date. For most solo devs and small teams that is four vendors and five minutes. The teams that skip it are not saving five minutes; they are choosing to discover pricing changes through their invoice, which is the most expensive notification channel on the market. We have been on both sides of that trade this month, and the cache billing episode taught us the same lesson from the infrastructure side: the vendors are not obligated to make noise when the economics of your setup change.
If your prompts already live in a library with notes and version history, this audit is a scan of one field. If they live in files and chat histories, this is the month that habit started costing real money: the full multi-vendor pricing picture has four vendors moving on four different axes, and each of them can move again next week. Record the price, date the record, and let the library remember what you checked and when. The prompt is yours; the price never was.
[2] Anthropic, Claude pricing documentation (updated August 2026): the 2/10 per million token introductory pricing for Claude Sonnet 5, originally scheduled to rise to 3/15 on September 1, 2026, is now the standard price; the change is attributed to the newer tokenizer producing roughly 30% more tokens for equivalent text. Announced August 11, 2026. https://platform.claude.com/docs/en/about-claude/pricing
