A single HTML file that tells you what one request — or a month of them — costs on 60 models, using a price table refreshed daily.
No build step, no dependencies, no tracking, no analytics. Open the file and it fetches one JSON price table over a CDN. Save it and it works offline against whatever the table said last.
One coding-agent request — 382,560 cached input + 16,280 fresh input + 1,160 output tokens, the measured 95.6% cache-hit mix — priced against the catalogue read 2026-11-17. The 12 cheapest of 99:
| $ per request | $ per month | Model | $ cache read | Clock |
|---|---|---|---|---|
| $0.0019 | $57.23 | DeepSeek V4.1 Flash 1017 batch |
$0.003 | flat |
| $0.003 | $90.50 | Claude Haiku 5.5 batch |
$0.005 | flat |
| $0.005 | $149.75 | DeepSeek V4.1 Flash 1017 | $0.01 | flat |
| $0.006 | $181.01 | Claude Haiku 5.5 | $0.01 | flat |
| $0.0095 | $284.05 | MiMo-V2.6-Pro | $0.0036 | flat |
| $0.0103 | $308.92 | GLM 5.2 | $0.006 | flat |
| $0.0107 | $320.20 | Nemotron 3.5 Super VL | $0.02 | flat |
| $0.0146 | $438.44 | Solar Pro 4 | $0.03 | flat |
| $0.0173 | $517.88 | Muse Glimmer 30B batch |
$0.032 | flat |
| $0.0199 | $596.77 | Step 3.7 Flash | $0.04 | flat |
| $0.0218 | $654.19 | MiniMax M3 batch |
$0.0448 | flat |
| $0.0226 | $678.78 | Gemini 3.6 Flash batch |
$0.0375 | flat |
$ per month is that request 1,000× a day for 30 days — change either number, or your own cache-hit share, and the page re-prices all 99 side by side: https://xyzs996.github.io/llm-cost-calculator/
The rows marked peak/off-peak bill by the clock, and the number above is the catalogue's single figure; the page applies whichever side is in force at the minute you ask, and tells you what waiting is worth. Long-context tiers are a cliff, not marginal pricing — the page re-rates the whole request the moment your prompt crosses the threshold.
A catalogue reports one price per model. That price is whichever seller is cheapest on the input column this minute — not the vendor's own rate. Open-weight models are resold by many hosts, and a cache read can cost one host several times what it costs another. 11 of the models in this table cannot be read as a single price (16 of the 99 rows, counting batch variants separately):
| Model | Sellers | Cache read, low → high | Sorting by input picks | Cheapest on the bill | Overpay |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017 | 50 | 2.0% → 37.5% | Wafer | Ionstream | +106.2% |
| Kimi K2.5 | 4 | 16.67% → 40.82% | AtlasCloud | Novita | +78.6% |
| Kimi K2.6 | 17 | 10.0% → 50.0% | Baidu | Chutes | +42.7% |
| Kimi K3.1 Pro | 24 | 5.0% → 83.23% | Relace | Decart | +22.5% |
| Kimi K3 | 27 | 10.0% → 99.35% | Wafer | Wafer | +21.2% |
| GLM 5.2 | 34 | 10.0% → 100.0% | Relace | Wafer | +15.5% |
The other 4 spread just as wide, but today the input column happens to land on the host that is also cheapest on the bill. That is luck, and it is re-drawn every time a seller reprices.
Overpay is what routing on the visible column costs you against routing on the bill, at the same cache-hit mix as the table above. Closed-weight models do not appear here: their resellers all bill a cache read at the vendor's mandated fraction of their own input price, so the row really is one price. Every figure comes from OpenRouter's per-model endpoint list and is refreshed with the rest of the table.
1. It applies DeepSeek's peak/off-peak rate for right now, not an average.
DeepSeek bills by the clock: peak on weekdays 01:00–04:00 and 06:00–10:00 UTC,
off-peak — half price — the rest of the time, and since 2026-08-22T16:00:00Z
all day on Beijing-time weekends. The page shows which side of that you are on,
what the same job costs on the other side, and how long until it flips. If you
are in a peak window it tells you what waiting is worth.
2. It treats the long-context tier as a cliff. For the models that have one, the moment your prompt reaches the threshold every token in that request re-rates — including the tokens below the threshold. It is not marginal pricing, and a calculator that bills the excess at the long rate understates it. The page warns you when you're near the edge and shows what crossing it costs.
3. Its default token mix is measured, not guessed. Coding agents are cache-heavy in a way that makes list prices misleading: the default here is 95.6% cache hit, 4.1% miss, 0.29% output, from 8.04B tokens of real agent traffic. The measurement is linked on the page. Change it to yours — the point is that the starting number came from somewhere.
Time-of-day billing has a failure mode that ordinary tests cannot catch. The Beijing weekend runs 16:00 UTC Friday to 16:00 UTC Sunday; the two calendars disagree only over 16:00–24:00 UTC; and both of DeepSeek's peak windows sit clear of that stretch. So an implementation that reads the weekday off the unshifted UTC instant passes every vector you can write against the live schedule, and starts mispricing by 2× the day a vendor moves a window later.
This page's phaseAt and nextChange are checked against a published,
dated vector table that includes a deliberately synthetic schedule built to pin
that axis:
node test/check.mjs # 18/18 passed
And the vectors have teeth — each failure mode injected into this page's own implementation is caught by the vectors that claim to catch it:
mutation to index.html |
vectors that fail |
|---|---|
| weekday read off the unshifted UTC instant | 2026-08-28T16:30:00Z, 2026-08-30T16:30:00Z |
| effective-date gate removed | 2 phase vectors + the pre-rule boundary vector |
| countdown does not skip weekend-resident edges | 2026-08-28T10:30:00Z |
The table lives at
xyzs996/deepseek-peak-offpeak-vectors
(CC0) and test/check.mjs downloads it, so the check runs against the published
copy rather than a vendored one that could drift.
Catalogue prices from the OpenRouter model list;
the models that bill by the clock are additionally checked against the vendor's
own pricing page, with the date of that check carried in the data. Both live in
xyzs996/llm-api-pricing as
data/prices.json
and a CSV beside it, refreshed daily. This repo holds no prices of its own — it
is a renderer, so it cannot go stale separately from the table.
Rounding, minimums, and unlisted surcharges are yours to verify. This is a calculator, not a quote.
Open an issue and say which one. A wrong number with the vendor page beside it is the most useful thing anyone can send.
The full price table, sortable, with the effective per-million cost under the measured agent mix: https://xyzs996.github.io/llm-api-pricing/prices.html. Write-ups on what these costs do to a small budget in practice — a different question from what the rate card says: https://xyzs996.github.io/llm-api-pricing/.
MIT.