Skip to content

Update DeepSeek V4 to the 2026-08-17 peak/off-peak prices (refs #70) - #71

Open
xyzs996 wants to merge 2 commits into
simonw:mainfrom
xyzs996:deepseek-peak-offpeak
Open

xyzs996 wants to merge 2 commits into
simonw:mainfrom
xyzs996:deepseek-peak-offpeak

Conversation

@xyzs996

@xyzs996 xyzs996 commented Aug 22, 2026

Copy link
Copy Markdown

Refs #70 — the numbers side of it, ready to merge or to throw away if you'd rather shape it differently.

Every DeepSeek V4 figure on the site predates the 2026-08-17 repricing. Source for the new ones: https://api-docs.deepseek.com/quick_start/pricing/, read 2026-08-22.

id input input_cached output
deepseek-v4-flash $0.14 → $0.22 $0.028 → $0.007 $0.28 → $0.66 renamed "off-peak"
deepseek-v4-flash-peak $0.44 $0.014 $1.32 new
deepseek-v4-pro $1.74 → $0.66 $0.145 → $0.022 $3.48 → $1.98 renamed "off-peak"
deepseek-v4-pro-peak $1.32 $0.044 $3.96 new

Peak is 01:00–04:00 and 06:00–10:00 UTC; everything else off-peak.

What this PR chooses, of the three options in #70: option 2, separate ids, because it reuses the -200k pattern the dataset already has and puts the 2x on the page instead of in a footnote. If you'd rather have option 1 (one row, multiplier in the name) or option 3 (a real time-window field), close this and I'll rewrite it — the figures above are what matters and they're the same either way.

Old prices get to_date: "2026-08-17" instead of being overwritten, so historical-v1.json keeps the pre-repricing record. python3 scripts/build.py runs clean and the DeepSeek block of current-v1.json comes out as the table above.

One thing these ids still can't express, flagged rather than hidden: from midnight Beijing time on 2026-08-23, DeepSeek bills Saturdays and Sundays at off-peak rates around the clock. from_date/to_date are calendar dates, so there's no day-of-week dimension either. The -peak rows are therefore an upper bound that's correct on weekdays and 2x too high at weekends. If that bothers you, option 1 collapses the problem into prose and might be the better trade after all.

I left deepseek-chat and deepseek-reasoner alone. They're gone from DeepSeek's pricing page entirely, so I have nothing live to quote for them — noted in #70 as possible to_date candidates rather than something to fill in.

Refs simonw#70. DeepSeek moved V4 Flash and V4 Pro to two-tier time-of-day
pricing effective 2026-08-17; every number in this file predates that.

Old prices get to_date 2026-08-17 rather than being overwritten, so
historical-v1.json keeps them. Peak rates go in as separate ids, following
the existing -200k convention for models that carry two price points.

Source: https://api-docs.deepseek.com/quick_start/pricing/ read 2026-08-22.
@xyzs996

xyzs996 commented Aug 23, 2026

Copy link
Copy Markdown
Author

Pushed one more commit: DeepSeek shipped deepseek-v4-flash-vision-exp on 2026-08-21 (release note), after I opened this. It's on the pricing page now and priced identically to V4-Flash on all six numbers, so it follows the same two-row shape as the rest of the PR, with from_date: "2026-08-21".

id input input_cached output
deepseek-v4-flash-vision-exp $0.22 $0.007 $0.66
deepseek-v4-flash-vision-exp-peak $0.44 $0.014 $1.32

python3 scripts/build.py still runs clean; current-v1.json goes from 154 to 156 models.

One thing the schema can't hold, flagged rather than folded in: images bill at up to 384 tokens each, so for image input these per-token figures are the multiplier, not the price.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant