dsh-meter English | 中文 DeepSeek bills by time of day now. This is the meter for it. Two peak windows a weekday, off-peak at half the peak rate — and since 22 August, off-peak all weekend. Cost stopped being a number you read afterwards and became a rate you are standing in. One line under the composer: what this session cost, which tariff is running, how long until it flips. Hover it for the tariff clock, the cache economics, and your balance.

Install

dsh plugin --profile web add @dshworks/dsh-meter
dsh --profile web

dsh plugin forwards to pnpm, so pnpm must be on PATH. Nothing else to configure: the meter appears under the composer as soon as a session bills its first request.

The line

2 turns · 2 steps | LLM 3.1s | TTFT avg 1.3s · 41 tok/s | Cache hit 50% | Input 44.3K tok
                    ¥0.0672  |  peak  |  off-peak in 48m

The harness's own stats line, untouched, and ours under it. This plugin adds a line — it does not shadow the shipped one, which is what a cost plugin has to do to append to it.

Hover or focus for the card. Inside a peak window the tariff is called out in amber and the countdown runs to the next off-peak hour:

The card shows	Why it is there
The session total, request count, models	The number, once, at full size
A 24-hour tariff strip in your local time, with a live now-marker	Peak windows are published in UTC and restricted to Beijing weekdays. Reading them off a strip beats doing timezone arithmetic at 11pm — and on a weekend the strip is empty, which is the answer
cache hits / fresh input / output — tokens and money on each	A cache hit costs 1/30th of a miss. This is the row that shows whether your prompt prefix is stable
Your account balance, and how much of it is granted credit	Granted balance expires; topped-up balance does not
The same tokens priced under the other tariff	Before the switchover: what the new rates do to this session. After: what waiting for off-peak is worth
What the cache saved	The counterfactual where every hit had been a miss

Settings: the meter before you spend

The line needs a session, and a session that has already billed something — so the one question this plugin exists to answer was unaskable until after you had paid to ask it. Settings → Meter answers it cold.

It is deliberately not a form. Both config knobs live in the plugin's own configuration, where the harness already edits plugin config; a second write path for two fields would be the settings-page reflex rather than the content. What is here is the instrument:

The section shows	Why it is there
The tariff running now, at full size, with the countdown and the clock time it ends	The reason to open the panel, answered before you read anything else
The week — seven local days by twenty-four hours, peak in amber, a live now-marker	The card's strip answers "when does today change". Since weekends went off-peak this is a weekly question, and 35 amber cells out of 168 is the fastest way to see that peak is 35 hours a week, not 49
The published rate card in the currency your account is billed in	At the precision DeepSeek publishes it — ¥1.5, not ¥1.50. A rate is a quotation, not a total
Your balance	Read once when the panel opens, through the same host route the card uses

The grid is laid out in your days, not Beijing's, and every cell asks the same function that prices a request — so the picture and the bill cannot disagree while still sparing you the timezone arithmetic.

In Chinese the hour slots carry a small joke. Peak and valley — 峰 and 谷 — are the old words for time-of-use electricity, and the peak half is one character from the name of the man who founded the company selling the tokens. Hover a slot: 文峰, 文谷.

Proof

Live-verified against dsh 0.1.1-rc.2 on 2026-08-24, in a real web session on DeepSeek-V4-Pro and V4-Flash — not a mock:

Re-run, because reading the source was not enough. Through rc.8 this note argued from a byte-identical src/ that the projection contract had not moved. It then moved: 0.1.1-rc.1 renamed schema to stateSchema and made the client-visible view an optional wire. Because the registry reads the definition field by field and never validates it, the old spelling did not fail — it silently meant host-only, and the dock line rendered nothing on an otherwise healthy harness. A contract that degrades to silence cannot be checked by diffing its source; the table below is a fresh live run, not an argument.

Claim	How it was checked
Loads in a stock web profile	dsh --profile web --dump-config lists @dshworks/dsh-meter; /plugins/@dshworks/dsh-meter/client.js serves 200
The line renders	¥0.2185 | off-peak | peak in 14h under the composer, read out of the live DOM — the check that would have caught the rc.1 rename
The readout is correct	36.1K cache-miss input + 147 output across V4-Flash and V4-Pro = ¥0.2162 + ¥0.0023 = ¥0.2185, matching the harness's own token counts on the stats line directly above it
Survives a restart	Server restarted, an eight-day-old session reopened cold — the projection replays from the durable log at the same figure
The card opens	Per-model split, the cache counterfactual (all-peak ¥0.2185 vs all-off-peak ¥0.1093), and the account balance, with no page errors
Currency detection	A live account returns {"currency":"cny", ...} from /dsh-meter/balance and the whole surface switches to ¥ with no configuration
Both themes, both tariff states	Light and dark, flat and peak, captured above on 2026-08-15 — that run predates the 08-16 switchover, so its flat readout is one a new session no longer reaches
130 tests, CI green	pnpm test — the fold, the tariff clock, the rate card, the saving-mode nudge, the balance reader, the registration contract against a stand-in registry of each era, and the generated-bundle sync check

Two currencies, no conversion

DeepSeek publishes two independent rate cards — USD for the international platform, CNY for the mainland one. An account is billed against exactly one of them, and neither table is a conversion of the other, so a plugin that picks with an exchange rate is wrong twice.

dsh-meter computes both and lets the account decide. GET /user/balance returns the currency the account is denominated in — a live account lists both rows with only one funded — so the meter reads the funded row and shows that side. Nothing to configure, nothing to guess from your interface language.

The balance request runs on the host, keyed by the same credential seam the LLM adapter uses; the browser gets parsed numbers over GET /dsh-meter/balance and never the key. It fires when the meter mounts and when you open the card, not on a schedule — set balance: false to turn the whole thing off and the meter keeps working.

The rate card

Carried verbatim in lib/core.js, per 1M tokens.

	cache hit	cache miss	output
v4-flash off-peak	$0.007 / ¥0.05	$0.22 / ¥1.5	$0.66 / ¥4.5
v4-flash peak	$0.014 / ¥0.10	$0.44 / ¥3	$1.32 / ¥9
v4-flash-vision-exp off-peak	$0.007 / ¥0.05	$0.22 / ¥1.5	$0.66 / ¥4.5
v4-flash-vision-exp peak	$0.014 / ¥0.10	$0.44 / ¥3	$1.32 / ¥9
v4-pro off-peak	$0.022 / ¥0.15	$0.66 / ¥4.5	$1.98 / ¥13.5
v4-pro peak	$0.044 / ¥0.30	$1.32 / ¥9	$3.96 / ¥27

Peak is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday (09:00–12:00 and 14:00–18:00 Beijing). Every other hour is off-peak, including the two-hour gap between the windows and the whole weekend. Off-peak is exactly half of peak — and still above the flat rate it replaced, by about 2.3x on output.

deepseek-v4-flash-vision-exp shipped 2026-08-21 and bills at exactly the v4-flash rates in both currencies. Images are converted to tokens by their dimensions and billed as input. It has no flat row: it arrived after the switchover, so it has no pre-time-of-use history to reprice.

Weekends have been off-peak all day since 2026-08-22 16:00 UTC (00:00 Beijing, Sunday 23 August), on the Beijing calendar — so the weekend turns over at 16:00 UTC, not at midnight UTC. Peak is 35 hours a week, not 49. DeepSeek announced this only in the pricing-page footnote and only until it took effect; the live page now states just the settled rule, so the announcement survives at the archived page.

The retired flat card, kept to reprice history

Billed at every hour until 2026-08-16 16:00 UTC. Upstream no longer publishes it; the meter keeps it because a session logged before the switchover must still cost what it actually cost.

	cache hit	cache miss	output
v4-flash flat	$0.0028 / ¥0.02	$0.14 / ¥1	$0.28 / ¥2
v4-pro flat	$0.003625 / ¥0.025	$0.435 / ¥3	$0.87 / ¥6

Against it, pro rose 6x/12x on cached input, 1.5x/3x on cache-miss input and 2.3x/4.6x on output (off-peak/peak). The steepest rise is on the cheapest token, which is the one an agent sends most of.

Source: https://api-docs.deepseek.com/quick_start/pricing, diffed against both locales daily (see below), last confirmed in sync 2026-08-20 — and against a real bill: 188,542 cache-miss tokens on pro, off-peak, settled at ¥0.84, i.e. ¥4.46/1M against the published 4.5.

The card is also a feed

Building your own estimator? Take the card, not the numbers in this table.

https://dsh.works/dsh-meter/pricing.json

Static JSON, no key, no rate limit. Both currencies, both tariffs, the weekly schedule, the retired flat card for repricing history, and the bucket definitions — generated from lib/core.js by scripts/build-feed.mjs, so it cannot state a price the meter would not charge.

In JavaScript, skip the fetch and call the same module directly:

import { costOf, tariffAt } from '@dshworks/dsh-meter/core'

const tokens = { miss: 188_542, hit: 1_204_880, out: 9_310 }
costOf(tokens, 'deepseek-v4-pro', tariffAt(Date.now()), 'cny')

lib/core.js imports nothing. Rate card, tariff clock, and cost fold are pure functions of their arguments — no clock reads inside, so history reprices at the tariff it was actually billed under.

What does it cost right now?

The feed answers this without containing a "now". It publishes the week as a grid; you index it with the current Beijing weekday and hour. That is why a CDN can cache it for ten minutes, or you can vendor it into a binary, and it still cannot be stale about which tariff is running:

curl -s https://dsh.works/dsh-meter/pricing.json | jq -r '
  ((now + 8*3600) | gmtime) as $b
  | .timeOfUse.scheduleBeijing[$b[6]][$b[3]] as $t
  | "\($t) · v4-pro out $\(.models["deepseek-v4-pro"].rates[$t].usd.out)/1M"'

Shift by 8*3600 first, then break down once, and take both indices off that same result: $b[6] is the day of week (0 = Sunday, as in JavaScript) and $b[3] the hour. An endpoint that returned the answer instead would be wrong for as long as its cache lives, on exactly the boundary where being wrong costs 2x.

Three ways to get this wrong, all silent:

• Local hours. scheduleBeijing is indexed on the vendor's clock and is not rotated into your timezone. new Date().getHours() returns a plausible tariff that is wrong for most of the planet.
• Mixing the two clocks. Taking the weekday from your own calendar and the hour from Beijing's disagrees for the eight hours from 16:00 UTC, when Beijing is already on the next day. Derive both from one shifted instant.
• Broken-down time offsets. jq's gmtime is [year, month, day, hour, minute, second, weekday, yearday] — the hour is .[3] and the weekday .[6]. Writing .[2] reads the day of the month, which is a valid index into a 24-hour array. Testing this section at 09:59 UTC, .[2] said offpeak while the real tariff was peak. It looked completely reasonable.

Reading @1's scheduleUtc now returns null: the key was removed rather than deprecated, because a reader that kept using it would have gone on labelling every weekend peak and overstating those sessions by 2x. See timeOfUse.changedInV2 in the feed.

The schedule is anchored to Beijing (UTC+8), and China has not observed daylight saving since 1991 — that fixed offset is the only reason a table of UTC hours is safe to publish at all. The feed states it under timeOfUse.anchor, and verify-pricing checks the English page's UTC sentence and the Chinese page's Beijing sentence separately, then checks that they still agree.

One caveat no feed can fix: now is your machine's clock. If it has drifted, so has your tariff.

Why not use an existing price feed? Because none of them are right. Checked 2026-08-20, four days after the switchover, both of the sources an estimator usually reaches for still published DeepSeek's retired flat card as current:

Source	v4-pro cache-miss input	Understates the real bill by
models.dev	$0.435	1.5x off-peak, 3.0x peak
LiteLLM	$0.435	1.5x off-peak, 3.0x peak
this feed	$0.66 off-peak / $1.32 peak	—

On cached input, the bucket an agent sends most of, models.dev is off by 12x at peak. And this is not a staleness bug they can patch: both schemas hold one flat price per model per bucket, with nowhere to put a tariff. A number that is wrong for seven hours of every weekday cannot be represented correctly in either.

OpenRouter's /api/v1/models is accurate, but for a different question — it quotes what OpenRouter charges to resell the model, not what DeepSeek bills your account.

Nobody types these numbers

Not us, and ideally not you. A hand-maintained rate card rots, and the vendor is the proof: DeepSeek's own pi integration guide ships a cost block where v4-flash's cache-read rate is a 10x decimal slip (0.028 for a rate of 0.0028) and the v4-pro figures are exactly 4x the card — they are Azure's resale prices, pasted into DeepSeek's documentation, on a page people copy into their agent config.

So the card is written by machine and reviewed by human:

read	scripts/verify-pricing.mjs	scrapes both locales — prices, peak windows, model list — and diffs them against the card
write	scripts/apply-pricing.mjs	splices the scraped values back into lib/core.js, then re-runs the verifier against its own output and reverts if that second read disagrees
ship	npm run build	regenerates the bundle, the site and the feed from the rewritten card
judge	you	the job opens a PR; it never merges one

A daily job runs all four. Locally it is npm run verify:pricing to look and npm run sync:pricing to rewrite.

Why it opens a PR instead of merging. Every invariant CI can check is structural — off-peak is half of peak, output costs more than a cache miss, a live request never prices at the retired card — and a wrong-but-plausible parse satisfies all of them. No test can tell a right price from a believable one. A human looking at a diff can.

The last price change arrived with no announcement in any channel we watch. That is the argument for the alarm; the vendor's own 10x typo is the argument for taking the keyboard away from all of us.

How the money is counted

Behavior	Detail
Source of truth	Provider-reported token counts in the durable session log. Nothing is sampled or inferred
Tariff per request	Decided by dispatch time (step/start), not by when the answer finished. A request sent at 00:59 UTC is billed off-peak even if it streams into the peak window, and a session that spans a boundary is billed correctly on both sides
Billed buckets	inputTokens at the cache-MISS rate, cacheReadTokens at the cache-HIT rate, outputTokens at the output rate. The harness reports these disjoint (the DeepSeek adapter subtracts hits out of prompt_tokens), and reasoning tokens are already inside output
Cache writes	Folded into the miss bucket. DeepSeek publishes no separate write price and bills a first-time prompt at the miss rate; the DeepSeek adapter never reports one
Failed requests	A step's usage chunk is counted even when the request then fails; the finalized message REPLACES that sample rather than adding to it, including when the message names a different model than the request header did
Unknown models	Counted and named, never priced. A model with no published rate contributes tokens and requests to the readout and zero money, and the card says so
Durability	One session projection (costMeter), folded from the log. It survives paging, compaction, a reload, and a server restart, and rides the standard projection cache

Model Experience

None by default. dsh-meter adds no tool, no system-prompt section, no message, and no model call; it does not touch the request. Cost belongs to the person paying, not to the agent's context window. The one opt-in exception is saving mode (savingMode: true): one system-prompt section that names the tariff in force, absent entirely when its text renders empty.

KV Cache effect

None by default. In saving mode, the meter:tariff section's text is byte-identical inside a tariff window, so the prompt prefix — and its cache — hold from one request to the next; the section changes only at the tariff boundaries, so the first request after a flip misses the prefix cache, and then holds again.

Configuration

Everything below is a validated config field, set in your profile's cordis.patch.yml:

- id: dsh-meter
  config:
    currency: cny        # pin a rate card instead of detecting it
    balance: false       # never call /user/balance
    savingMode: true     # tell the model which tariff it is standing in

Key	Default	Meaning
currency	auto	Which rate card to show: auto (the account's own balance currency, then the interface language), usd, or cny
balance	true	Serve the account balance to the Web UI
apiKeyEnv	DEEPSEEK_API_KEY	Credential reference holding the API key
baseUrl	https://api.deepseek.com	Origin the balance is read from
balanceTtlMs	300000	Minimum age before a card opening refetches the balance
balanceTimeoutMs	4000	Per-request timeout for the balance read
savingMode	false	Contribute a system-prompt section naming the tariff in force, so the model can behave accordingly
savingPeakPrompt	the built-in nudge	Text injected while a peak window runs
savingOffPeakPrompt	''	Text injected while no peak window runs; empty (the default) means the section renders to nothing and costs zero prompt tokens

Saving mode — the meter talks back

Off by default, because a readout should never grow into the agent's context window on its own. Flip savingMode: true and the meter stops being a bystander: at every assembly it contributes one system-prompt section, meter:tariff, that tells the model which tariff the next request will be dispatched under and how to behave in it.

Inside a peak window the section carries savingPeakPrompt — by default the meter's own nudge: the peak hours are named, the doubling of the bill is stated, and the model is asked to answer concisely, prefer context it already has over new tool calls, and offer to defer expensive work until the window closes. Outside the windows the section renders savingOffPeakPrompt; empty by default, so the meter's voice costs zero tokens when there is nothing to warn about. Want the model to loosen up off-peak instead? Set savingOffPeakPrompt to something like "off-peak, the half-price hours are running — you may be liberal with tokens".

Four properties are deliberate:

• The tariff is read at assembly time, from the same clock the billing fold uses. The nudge and the bill can never disagree about which side of a boundary the next request lands on — including weekends, which bill off-peak all day and therefore get no nudge at all.
• The hours in the nudge are derived, not typed. peakHoursPhrase() reads them off PEAK_WINDOWS_UTC and tariffSchedule(), so a schedule change rewrites the sentence. A prompt that tells a model the wrong peak hours is a plugin lying to it about money.
• The text is constant inside a tariff window. A countdown in here would change every minute and roll the session's prompt-prefix cache for one sentence's sake — the most expensive way to save money this plugin can think of. The section flips only at tariff boundaries (first request after a flip misses the prefix cache), and then holds.
• It is a nudge, not a throttle. The model may ignore it; nothing is enforced at the request layer. If you need a hard cap, that is a different plugin.

Development

pnpm install
pnpm test                        # generated-artifact checks, then vitest
pnpm run build                   # regenerate all three artifacts after editing core/ui
pnpm run verify:pricing          # diff the card against DeepSeek's own pages (network)

lib/core.js holds the rate card, the tariff clock, and the fold. lib/balance.js is the account reader. src/ui.js is the browser surface. Three artifacts are generated from that one card — lib/client.js (the bundle the harness serves), docs/index.html (the site) and docs/pricing.json (the feed) — so the price table exists in exactly one file, and pnpm test fails if any of them drifts from it.

verify:pricing is the only script that touches the network, and it is deliberately outside pnpm test: a unit suite must not fail because a documentation site is slow, and a price alarm must not stay silent because nobody opened a PR this week. It runs on its own daily schedule.

Known limitations

• It is an estimate, not the invoice. List prices times reported tokens. Promotional pricing and any relay in front of the API are invisible to it.
• Tariff is inferred from the dispatch timestamp, which is the harness's clock. A request DeepSeek receives on the far side of a boundary can bill differently by a second or two.
• Non-DeepSeek routes are not priced. Their tokens are counted and their model named; the card reports them as unpriced instead of quietly applying DeepSeek's rates to someone else's API.
• The rate card is compiled in. DeepSeek changes prices; this plugin ships the time-of-use card as re-checked on 2026-08-17 and needs a release to follow the next change.
• Web only. The projection is available to any surface, but the readout is built for the Web UI. There is no TUI line.

Prior art

The dsh registry lists ~50 cost and usage plugins; several arrived the same week DeepSeek dated the switchover, and reading them shaped this one. The cache-savings framing comes from deepseek-cli's local usage ledger.

License

MIT. Not affiliated with DeepSeek. "DeepSeek Harness" is DeepSeek's trademark, used here only to say what this works with; the name follows the "DSH" form their brand guidelines recommend.