the repricing is live — apply it everywhere, and check it against a bill - #22
Merged
Conversation
… it against a bill DeepSeek's peak/off-peak card took effect at 16:00 UTC on 2026-08-16. The switch itself was already encoded and dated, so cost estimates flipped on their own; what had not moved was every number and sentence written while the flat card was still real. Verified four ways before changing anything: the official EN and zh-cn pages (raw, not summarised) now show only the peak/off-peak table and the "will take effect" footnote is gone; our docs mirror's history supplies the old baseline; the widely-reported Chinese figures reconcile against it exactly (+203% / +355% / +1114% on pro); and one 188,542-cache-miss-token pro call settled at 0.84 CNY off-peak, i.e. 4.46 CNY/1M against the published 4.5 and the old card's 3.0. The billing system flipped, not just the page. - Site: the peak/off-peak card now leads on /pricing/ and /cost/, with the flat card kept below as what the ledger reprices history under. New lead news entry. The no-JS strip says billing is time-of-day rather than announcing a future date. - Corrected everywhere: the cache discount is ~30x, not 50x/120x, and the kill-line ratios on /bench/ are 7.6x/15x/23x off-peak, not 11.5x/34.5x/138x. The 138x cache-read edge was the load-bearing number in several field reports, so those now carry a dated caveat. - The playground priced calls off the dead card and ignored the clock; it now reads both, like the CLI. - gateway/DESIGN.md's worst-case free-tier arithmetic doubled: $0.0264 per user per day off-peak, ~38 users per $1 instead of ~70. - Two tests were flip-day time bombs, asserting today's price against literals from the old card. Repriced at fixed instants: what they test is token attribution, not which card is in force. - make corpus now excludes en/dsh. The mirror started tracking the harness docs, which would have quadrupled the embedded payload and outranked the API pages on shared words — it already broke a search ranking test. Corpus is back to 68 pages, 84K, with the new pricing. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
DeepSeek's peak/off-peak card took effect at 16:00 UTC on 2026-08-16. The switch was already encoded and gated on its date, so cost estimates flipped by themselves — what had not moved was every number and sentence written while the flat card was still real.
Verified before changing anything
The last one is the point: a docs page can update before a billing system does. It didn't.
What changed
/pricing/and/cost/; the flat card stays below as what the ledger reprices history under. New lead news entry. The no-JS strip states time-of-day billing instead of announcing a future date./bench/are 7.6x / 15x / 23x off-peak, not 11.5x / 34.5x / 138x. That 138x cache-read edge was load-bearing in several quoted field reports, so those carry a dated caveat now rather than being silently rewritten.gateway/DESIGN.md— worst-case free-tier cost doubled: $0.0264/user/day off-peak, ~38 users per $1 instead of ~70. The meter code was already correct.make corpusnow excludesen/dsh. The mirror started tracking the harness docs (112 pages); packing them quadrupled the embedded payload and outranked the API pages on shared words, breaking a search-ranking test. Back to 68 pages / 84K, with the new pricing.make checkpasses. Pages checked at 1280 / 768 / 420 / 360 — no sideways scroll, and the news table was cut from 5 columns to 3 so a phone shows a whole model rather than half of one.🤖 Generated with Claude Code