diff --git a/AGENTS.md b/AGENTS.md index ff4db72..9dc6bac 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -105,10 +105,11 @@ deepseek usage --entries --json # individual calls deepseek pricing --json # the schedule and the billing period right now ``` -Pricing is time-of-day from 2026-08-16 16:00 UTC: peak hours 01:00–04:00 -and 06:00–10:00 UTC bill at twice the off-peak rate. `pricing` computes -the current period locally — no network, nothing spent — from the same -schedule the cost estimates use. +Pricing has been time-of-day since 2026-08-16 16:00 UTC: peak hours +01:00–04:00 and 06:00–10:00 UTC bill at twice the off-peak rate. Never +quote a DeepSeek price without saying which period it is — `pricing` +computes the current one locally, no network, nothing spent, from the +same schedule the cost estimates use. ### Documentation, offline @@ -264,9 +265,11 @@ Every call prints a usage line to stderr and appends to for a quiet run. - Costs are estimates from the published USD rate card, not billed amounts. Token counts are exact. -- Cached input tokens cost 50× less than uncached ones. Put the stable - part of a prompt first — the same system prompt and files across calls - — and `deepseek usage` will show the saving. +- Cached input tokens cost about 30× less than uncached ones. Put the + stable part of a prompt first — the same system prompt and files across + calls — and `deepseek usage` will show the saving. Prompt structure is + still the biggest lever on a bill; the hour of the day comes second, + and it is worth at most 2×. ## Cautions diff --git a/Makefile b/Makefile index 5e5137a..98a7e9d 100644 --- a/Makefile +++ b/Makefile @@ -32,19 +32,25 @@ cover-gate: # page by page and extracts the FAQ out of its JS bundle. A sibling # checkout is used when there is one — that is the loop while working on # both — and otherwise it is fetched, so CI and a bare clone both work. +# +# The mirror also tracks dsh, DeepSeek's agent harness, in en/dsh — 112 +# pages of a different product. This binary carries the API reference, so +# that subtree is excluded: it would quadruple the payload and it out-ranks +# the API pages on shared words like "tool" and "session". CORPUS=internal/docs/corpus.tar.gz +CORPUS_EXCLUDE=--exclude=en/dsh DOCS_REPO=https://github.com/thevibeworks/deepseek-docs .PHONY: corpus corpus: @if [ -d ../deepseek-docs/content/en ]; then \ echo "packing from ../deepseek-docs"; \ - tar -C ../deepseek-docs/content -czf $(CORPUS) en; \ + tar -C ../deepseek-docs/content $(CORPUS_EXCLUDE) -czf $(CORPUS) en; \ else \ echo "fetching from $(DOCS_REPO)"; \ tmp=$$(mktemp -d); \ curl -sSL $(DOCS_REPO)/archive/refs/heads/main.tar.gz | tar -xz -C $$tmp --strip-components=1; \ - tar -C $$tmp/content -czf $(CORPUS) en; \ + tar -C $$tmp/content $(CORPUS_EXCLUDE) -czf $(CORPUS) en; \ rm -rf $$tmp; \ fi @echo "$(CORPUS): $$(du -h $(CORPUS) | cut -f1), $$(tar tzf $(CORPUS) | grep -c '\.md$$') pages" diff --git a/README.md b/README.md index c113392..d11bbc3 100644 --- a/README.md +++ b/README.md @@ -54,9 +54,11 @@ It also does three things `curl` will not: on every later request. `--continue` handles that; without tools it strips the same field, because replaying it there just burns tokens. - **Prices every call.** DeepSeek's disk KV-cache makes a cached input - token **50× cheaper** than an uncached one. That split is invisible - unless something reads `prompt_cache_hit_tokens` and does the - arithmetic. This does, on every call, and keeps a local ledger. + token **~30× cheaper** than an uncached one, and since 2026-08-16 the + hour of the day doubles the bill on top. Both splits are invisible + unless something reads `prompt_cache_hit_tokens`, knows the schedule + and does the arithmetic. This does, on every call, and keeps a local + ledger. - **Carries the manual.** `deepseek docs ask "..."` answers questions about the DeepSeek API from DeepSeek's own documentation, offline, with a citation per claim. The tool that talks to an API should be able to @@ -374,9 +376,10 @@ context cache saved ~$0.23 (1.7M of 2.2M prompt tokens replayed) costs are estimates from the published USD rate card, not billed amounts ``` -That last line is the point. Cached input costs $0.0028/M against -$0.14/M for a miss — structuring prompts so the stable part comes first -is worth real money, and this is how you see whether it worked. +That last line is the point. Cached input costs $0.007/M against +$0.22/M for a miss in an off-peak hour — structuring prompts so the +stable part comes first is worth real money, and this is how you see +whether it worked. Honest limits: diff --git a/gateway/DESIGN.md b/gateway/DESIGN.md index bd8b429..ef9f854 100644 --- a/gateway/DESIGN.md +++ b/gateway/DESIGN.md @@ -280,20 +280,29 @@ scarce operation and gets the conservative boundary. ## Limits, and why these numbers -Worst-case cost per anonymous user per day, at flash rates -($0.14/M input, $0.28/M output): +Worst-case cost per anonymous user per day, at flash off-peak rates +($0.22/M input, $0.66/M output): ``` -60,000 input × $0.14/M = $0.0084 -20,000 output × $0.28/M = $0.0056 +60,000 input × $0.22/M = $0.0132 +20,000 output × $0.66/M = $0.0132 ------- - $0.014 / user / day at full burn + $0.0264 / user / day at full burn, off-peak + $0.0528 / user / day at full burn, peak ``` -So a $1/day budget serves ~70 users burning *everything*, or several -hundred normal ones — a normal turn is a few hundred input and a few -hundred output tokens, about $0.0003. Roughly **3,300 ordinary turns per -dollar.** +So a $1/day budget serves ~38 users burning *everything* off-peak, ~19 at +peak, or several hundred normal ones — a normal turn is a few hundred +input and a few hundred output tokens, about $0.0006. Roughly **1,700 +ordinary turns per dollar off-peak, 850 at peak.** + +These are the numbers of the card that took effect 2026-08-16 16:00 UTC, +and they roughly halved the free tier's reach at an unchanged budget: the +old flat card ($0.14/$0.28) made the same worst case $0.014/user/day. The +quotas below have not been retuned for it — the budget breaker is what +actually bounds spend, and it does its job at any card — but the reach +per dollar is now a time-of-day figure, and the peak windows are the +expensive seven hours. | setting | default | why | |---|---|---| diff --git a/gateway/internal/meter/meter_test.go b/gateway/internal/meter/meter_test.go index 77cfb1a..7013667 100644 --- a/gateway/internal/meter/meter_test.go +++ b/gateway/internal/meter/meter_test.go @@ -7,6 +7,12 @@ import ( "time" ) +// A fixed instant outside every peak window, for the tests that are about +// token attribution rather than about the schedule. Pricing them at +// time.Now() made them fail the moment the repricing landed, which told +// us nothing about the code under test. +var offPeakInstant = time.Date(2026, 8, 17, 12, 0, 0, 0, time.UTC) + // The payloads below are verbatim from the live API on 2026-08-05, one // per wire format, captured with `deepseek raw`. Hand-written fixtures // would only prove this package agrees with my memory of the shapes; @@ -69,9 +75,10 @@ func TestAnthropicCacheReadsAreAddedToInput(t *testing.T) { t.Errorf("cache hits = %d, want 900", u.CacheHitTokens) } // And the price must reflect that only 130 tokens were billed at the - // full rate. - want := 900*0.0028/1e6 + 130*0.14/1e6 + 7*0.28/1e6 - if got := Cost("deepseek-v4-flash", u); math.Abs(got-want) > 1e-12 { + // full rate. Priced at a fixed off-peak instant: what is under test is + // which tokens land on which rate, not which card is in force today. + want := 900*0.007/1e6 + 130*0.22/1e6 + 7*0.66/1e6 + if got := CostAt("deepseek-v4-flash", u, offPeakInstant); math.Abs(got-want) > 1e-12 { t.Errorf("cost = %v, want %v", got, want) } } @@ -82,8 +89,8 @@ func TestOpenAICacheHitsAreInsideInput(t *testing.T) { if u.InputTokens != 1000 || u.CacheHitTokens != 960 { t.Fatalf("in %d hit %d, want 1000/960", u.InputTokens, u.CacheHitTokens) } - want := 960*0.0028/1e6 + 40*0.14/1e6 + 10*0.28/1e6 - if got := Cost("deepseek-v4-flash", u); math.Abs(got-want) > 1e-12 { + want := 960*0.007/1e6 + 40*0.22/1e6 + 10*0.66/1e6 + if got := CostAt("deepseek-v4-flash", u, offPeakInstant); math.Abs(got-want) > 1e-12 { t.Errorf("cost = %v, want %v", got, want) } } diff --git a/internal/cli/e2e_test.go b/internal/cli/e2e_test.go index a1c5823..3dee18a 100644 --- a/internal/cli/e2e_test.go +++ b/internal/cli/e2e_test.go @@ -412,7 +412,11 @@ func TestModelsJoinsThePublishedRateCard(t *testing.T) { if !strings.Contains(got.stdout, "deepseek-v4-flash") { t.Errorf("stdout = %q", got.stdout) } - if !strings.Contains(got.stdout, "0.14") { + // Which figure depends on the hour: flash cache-miss input is $0.22 + // off-peak and $0.44 peak. Both are published numbers, so asserting + // on either keeps this a real check without making it a time bomb + // that fails whenever the suite runs inside a peak window. + if !strings.Contains(got.stdout, "0.22") && !strings.Contains(got.stdout, "0.44") { t.Errorf("the price should sit next to the model, got %q", got.stdout) } } diff --git a/internal/deepseek/pricing.go b/internal/deepseek/pricing.go index 42057c3..c162ed0 100644 --- a/internal/deepseek/pricing.go +++ b/internal/deepseek/pricing.go @@ -37,11 +37,16 @@ type Price struct { Output float64 } -// RepriceAt is when DeepSeek's dated repricing takes effect: 16:00 UTC -// on 2026-08-16 (midnight, Beijing), announced 2026-08-13 with the V4 GA +// RepriceAt is when DeepSeek's repricing took effect: 16:00 UTC on +// 2026-08-16 (midnight, Beijing), announced 2026-08-13 with the V4 GA // release. From that instant the API bills peak/off-peak on a new, // higher card, with off-peak at half the peak rate. // +// This is live, and measured, not just read off the page: a 188,542 +// cache-miss-token call to pro in the off-peak window on 2026-08-17 +// settled at 0.84 CNY, i.e. 4.46 CNY/1M against the new card's 4.5 and +// the old card's 3.0. +// // TASTE.md's rule against applying announced-but-undated numbers does // not apply here — these numbers carry their date, so the switch is // encoded and gated on it, exactly as that scar's expiry clause says. @@ -49,6 +54,10 @@ type Price struct { var RepriceAt = time.Date(2026, time.August, 16, 16, 0, 0, 0, time.UTC) // pricesFlat is the card published 2026-08-02, in force before RepriceAt. +// That instant has passed, so nothing live prices against it any more; it +// stays because the ledger stores token counts rather than dollars, and a +// call made before the flip must still reprice under the card it was +// actually billed at. var pricesFlat = map[string]Price{ ModelFlash: {CacheHitInput: 0.0028, CacheMissInput: 0.14, Output: 0.28}, ModelPro: {CacheHitInput: 0.003625, CacheMissInput: 0.435, Output: 0.87}, diff --git a/internal/docs/corpus.tar.gz b/internal/docs/corpus.tar.gz index f205e73..b6560f2 100644 Binary files a/internal/docs/corpus.tar.gz and b/internal/docs/corpus.tar.gz differ diff --git a/site/bench/index.html b/site/bench/index.html index 05a4d70..64cf120 100644 --- a/site/bench/index.html +++ b/site/bench/index.html @@ -30,7 +30,7 @@ """, @@ -1423,7 +1462,7 @@ def jstr(s): ("What did the V4-Pro GA (0813) checkpoint change over the preview?", "The model ID and the rate card did not change; the checkpoint did. Against the April V4-Pro preview, DeepSeek's chart shows large agentic gains: DeepSWE 12.8 to 62.7, DSBench-Hard 31.1 to 67.2, CyberGym 52.7 to 83.3, Terminal-Bench 2.1 72.1 to 87.9, Toolathlon 55.9 to 74.1. Jumps that size point to agent post-training and better tool-error handling rather than a new base model, and none of them are independently verified yet."), ("What is the DeepSeek kill line (斩杀线)?", - "The kill line is a community idea that DeepSeek's price-to-capability ratio sets a threshold that removes the reason to exist for any model that is both weaker and more expensive. V4-Pro is roughly 11x cheaper than GPT-5.6 Sol on cache-miss input and 34x cheaper on output, and about 138x cheaper on cache-hit input. It does not kill the frontier: the strongest closed models still finish the hardest tasks in fewer turns. It kills the middle, where a model costs more and does less."), + "The kill line is a community idea that DeepSeek's price-to-capability ratio sets a threshold that removes the reason to exist for any model that is both weaker and more expensive. On the card in force since 2026-08-16, V4-Pro is 7.6x cheaper than GPT-5.6 Sol on cache-miss input off-peak (3.8x at peak), 15x cheaper on output (7.6x peak) and 23x cheaper on cache-hit input (11x peak). The repricing narrowed all three: on the flat card that ran until then the figures were 11.5x, 34.5x and 138x. It does not kill the frontier: the strongest closed models still finish the hardest tasks in fewer turns. It kills the middle, where a model costs more and does less."), ("Is DeepSeek-V4-Pro better than Claude or GPT for coding agents?", "Per attempt, the strongest closed models remain more reliable on the hardest multi-step tasks and usually need less steering. Per dollar, V4-Pro changes the arithmetic: its cheap cached input makes repeated review, parallel workers and long tool loops affordable in a way per-token-stronger models are not. The practical answer is to route by role, run an internal bake-off, and measure successful-task cost, not per-token price."), ]), @@ -1530,20 +1569,27 @@ def jstr(s): - - - + + +
Per 1M tokensv4-flashv4-proGPT-5.6 Solpro is cheaper by
input, cache miss$0.14$0.435$5.00~11.5x
input, cache hit$0.0028$0.003625$0.50~138x
output$0.28$0.87$30.00~34.5x
input, cache miss$0.22 / $0.44$0.66 / $1.32$5.007.6x / 3.8x
input, cache hit$0.007 / $0.014$0.022 / $0.044$0.5023x / 11x
output$0.66 / $1.32$1.98 / $3.96$30.0015x / 7.6x
-

DeepSeek prices are the published USD rate card of -2026-08-02, unchanged at GA; they are a conversion of the RMB card -(¥3 / ¥0.025 / ¥6 per 1M for pro) at one consistent rate. GPT-5.6 -Sol prices are from OpenAI's own listing. From 2026-08-16 16:00 UTC DeepSeek -bills peak/off-peak on a higher card – the -pricing page has the dated schedule. Even at -the new peak rate, pro stays ~3.8x cheaper than GPT-5.6 Sol on cache-miss -input and ~7.6x on output.

+

DeepSeek cells read off-peak / peak, on the +card in force since 2026-08-16 16:00 UTC; they are a conversion of the RMB +card (pro: ¥4.5 / ¥0.15 / ¥13.5 per 1M off-peak, double at peak) +at one consistent rate. GPT-5.6 Sol prices are from OpenAI's own listing. +The pricing page has the full schedule.

+
+the repricing moved this line +

These ratios were 11.5x / 138x / 34.5x on the flat card +that ran until 2026-08-16. The cache-hit column is where the argument +lived – a 138x edge is what made replayed context, parallel reviewers +and long tool loops nearly free – and it is now 23x off-peak, 11x at +peak. That is still a large advantage. It is no longer a different +category, and any plan that was built on the old number should be +re-costed rather than assumed.

+

The sober version matters as much as the slogan. The kill line is real for the middle of the market: a model that costs more than V4-Pro and scores below it on the table above is hard to justify, and that is most of the @@ -1564,11 +1610,16 @@ def jstr(s): ds chat -m deepseek-v4-pro and the Anthropic remap make the switch one flag.

  • Structure prompts for the cache. A cached input token -costs about 1/50th of an uncached one. Keep the system prompt, tool schemas, +costs about 1/30th of an uncached one. Keep the system prompt, tool schemas, repository map and durable instructions in an identical prefix and put the volatile part last; ds usage reports what the cache saved so you -can see whether it is working. This is where the 138x cache-hit number turns -from a table cell into a bill.
    • +can see whether it is working. This is where the cache-hit column turns from +a table cell into a bill – and since the repricing hit that column +hardest, it is worth more attention now, not less. +
  • Schedule what can be scheduled. The same call costs +half as much outside 01:00–04:00 and 06:00–10:00 UTC. Batch +evaluation, bulk review and overnight agent runs are exactly the workloads +that can move; ds pricing says which period you are in.
  • Measure successful-task cost, not per-token price. A cheaper model that retries five times can cost more than a dearer one that lands first. The ledger stores exact token counts @@ -1591,9 +1642,9 @@ def jstr(s): PAGES.append(dict( slug="news/", crumb="news", - title="DeepSeek API news: the dated repricing, dsh ships, V4-Pro GA", - description="What is changing around the DeepSeek API: the price rise now has a date and numbers (peak/off-peak billing from 2026-08-16 16:00 UTC), DeepSeek ships dsh (DeepSeek Harness), its official open-source agent harness, V4-Pro's official release (DeepSeek-V4-Pro-0813), V4-Flash's official release, and what each one means for a call.", - keywords="deepseek harness, dsh, @deepseek-ai/dsh, install dsh, dsh plugins, dsh skills, dsh-plugin, dsh vs claude code, deepseek cli vs dsh, deepseek v4 pro release, deepseek v4 pro ga, deepseek-v4-pro-0813, deepseek api price increase, deepseek price rise 2026, deepseek peak hour pricing, deepseek peak off-peak billing, deepseek repricing 2026-08-16, deepseek api news, deepseek api changelog, deepseek v4 flash release, deepseek pricing change", + title="DeepSeek API news: the repricing is live, dsh ships, V4-Pro GA", + description="What is changing around the DeepSeek API: peak/off-peak billing went live at 2026-08-16 16:00 UTC and is confirmed against a real bill, DeepSeek ships dsh (DeepSeek Harness), its official open-source agent harness, V4-Pro's official release (DeepSeek-V4-Pro-0813), V4-Flash's official release, and what each one means for a call.", + keywords="deepseek harness, dsh, @deepseek-ai/dsh, install dsh, dsh plugins, dsh skills, dsh-plugin, dsh vs claude code, deepseek cli vs dsh, deepseek v4 pro release, deepseek v4 pro ga, deepseek-v4-pro-0813, deepseek api price increase, deepseek price rise 2026, deepseek peak hour pricing, deepseek peak off-peak billing, deepseek repricing 2026-08-16, deepseek new prices live, deepseek api news, deepseek api changelog, deepseek v4 flash release, deepseek pricing change", jsonld=faq([ ("Is DeepSeek V4-Pro officially released?", "Yes. On 2026-08-12 the model version on DeepSeek's Models & Pricing page changed to DeepSeek-V4-Pro-0813, ending the preview that had run since 2026-04-24. The model ID is unchanged (deepseek-v4-pro), the rate card is unchanged, and the release focuses on agentic post-training rather than new pretraining."), @@ -1601,12 +1652,12 @@ def jstr(s): "dsh is DeepSeek's official open-source agent harness, released 2026-08-13 as deepseek-ai/deepseek-harness on GitHub and @deepseek-ai/dsh on npm (MIT). It is in developer preview with compatibility-breaking changes promised. Everything in it is a plugin, built on the Cordis runtime; `npx @deepseek-ai/dsh web` starts its Web UI. The official repository does not accept external pull requests, and there is no plugin marketplace or registry: discovery is the dsh-plugin GitHub topic, with the ecosystem explicitly delegated to the community."), ("Is dsh the same as deepseek-cli?", "No. dsh (DeepSeek Harness) is DeepSeek's official agent harness: it runs an agent loop, executes tools, and manages plugins, skills and sessions. deepseek-cli is an unofficial single-binary API client: it sends one request in any of DeepSeek's four wire formats, prints the response and its estimated cost, and carries DeepSeek's API documentation offline. They are complementary, not competing: run agents with dsh, and use deepseek-cli to check a key, price a call, debug the wire formats, or query the docs."), - ("Is DeepSeek raising its API prices?", - "Yes, and the rise now has a date and numbers. At 16:00 UTC on 2026-08-16 DeepSeek moves to peak/off-peak billing on a new, higher card: peak hours are 01:00-04:00 and 06:00-10:00 UTC daily at twice the off-peak rate, and all other hours are off-peak. Per 1M tokens (cache hit / miss / output), deepseek-v4-flash goes to $0.007/$0.22/$0.66 off-peak and $0.014/$0.44/$1.32 peak; deepseek-v4-pro to $0.022/$0.66/$1.98 off-peak and $0.044/$1.32/$3.96 peak. This resolves both the undated broad price rise announced on 2026-08-06 and the undated peak-hour policy announced in June 2026."), - ("When does DeepSeek's peak-hour pricing start?", - "At 16:00 UTC on August 16, 2026. Peak hours are 01:00-04:00 and 06:00-10:00 UTC (09:00-12:00 and 14:00-18:00 Beijing time) daily, at twice the off-peak rate; every other hour is off-peak. Before that instant the flat card of 2026-08-02 applies at every hour and no multiplier should be applied."), + ("Did DeepSeek raise its API prices?", + "Yes. The increase took effect at 16:00 UTC on 2026-08-16 and is live: DeepSeek now bills peak/off-peak, with peak hours 01:00-04:00 and 06:00-10:00 UTC daily at twice the off-peak rate. The flat card is gone from the official pricing page. Against it, deepseek-v4-pro peak is 3x on cache-miss input, 4.6x on output and 12x on cache-hit input; off-peak is half of each. This resolved both the undated broad price rise announced on 2026-08-06 and the undated peak-hour policy announced in June 2026."), ("What are DeepSeek's current API prices?", - "Per 1M tokens, from the rate card published 2026-08-02 and in force until 16:00 UTC on 2026-08-16: deepseek-v4-flash is $0.14 input on a cache miss, $0.0028 on a cache hit, and $0.28 output; deepseek-v4-pro is $0.435 input on a miss, $0.003625 on a hit, and $0.87 output. From that instant peak/off-peak billing applies on a new card."), + "Per 1M tokens (cache hit / cache miss / output), on the card in force since 16:00 UTC on 2026-08-16: deepseek-v4-flash is $0.007 / $0.22 / $0.66 off-peak and $0.014 / $0.44 / $1.32 peak; deepseek-v4-pro is $0.022 / $0.66 / $1.98 off-peak and $0.044 / $1.32 / $3.96 peak. In RMB, pro is 0.15 / 4.5 / 13.5 yuan off-peak and 0.3 / 9 / 27 yuan peak. Peak hours are 01:00-04:00 and 06:00-10:00 UTC; every other hour is off-peak."), + ("Which hours are cheapest on the DeepSeek API?", + "Every hour outside 01:00-04:00 and 06:00-10:00 UTC, which bill at half the peak rate. Those peak windows are 09:00-12:00 and 14:00-18:00 Beijing time, the Chinese working day, so the whole European and American working day is off-peak. Batch and agent work that can be scheduled should run off-peak; the saving is exactly 2x."), ("Where can I follow DeepSeek API changes?", "DeepSeek's own change log lives at api-docs.deepseek.com/updates. The deepseek CLI carries the same documentation inside the binary: `deepseek docs changelog` prints it offline, and `deepseek docs sync` refreshes the snapshot."), ]), @@ -1617,7 +1668,83 @@ def jstr(s): live API where that is possible; the in-terminal feed is ds docs changelog.

    -

    2026-08-13 · the price rise has its date and its numberseffective 2026-08-16

    +

    2026-08-16 · the repricing is liveconfirmed against a bill

    +

    It landed on schedule. At 16:00 UTC on 2026-08-16 – +midnight in Beijing – DeepSeek's peak/off-peak card took effect, and +the Models & Pricing page no +longer carries the flat card at all: the old figures are deleted, and the +footnote that used to say the new prices “take effect at 16:00 UTC on +August 16” now simply describes how billing works. Both the English +and Chinese editions read the same way.

    + +
    + + + + + + + +
    Per 1M tokensv4-flashv4-pro
    input, cache hit$0.007 / $0.014$0.022 / $0.044
    input, cache miss$0.22 / $0.44$0.66 / $1.32
    output$0.66 / $1.32$1.98 / $3.96
    +
    +

    Cells read off-peak / peak. Peak hours are +01:00–04:00 and 06:00–10:00 UTC daily – 09:00–12:00 +and 14:00–18:00 Beijing, seven hours a day – and every other hour +is off-peak at half the peak rate. In RMB, pro is +¥0.15 / ¥4.5 / ¥13.5 off-peak and ¥0.3 / ¥9 / ¥27 at +peak.

    +

    Against the flat card of 2026-08-02, off-peak / peak: +pro is 6x / 12x on cached input, 1.5x / 3x on cache-miss +input and 2.3x / 4.6x on output. Flash is 2.5x / 5x, 1.6x / 3.1x and +2.4x / 4.7x.

    + +

    Read the increase where it actually bites

    +

    The headline “up to 3x” is the cache-miss number, and it is the +least interesting one. The steep rise is on pro's cached input: 12x +at peak, 6x off-peak. That is the token DeepSeek was famous for +pricing at almost nothing, and it is the token an agent sends most of – +every replayed system prompt, every re-sent file, every turn of a long tool +loop. A pro workload built on a 99%-cache-hit prefix does not see a 3x bill; +it sees something much closer to 12x. The cheapest thing on the menu went up +the most.

    +

    Note that this one is pro-specific. Flash's cached input rose 2.5x +off-peak and 5x at peak, so the two models' cache rates have converged: +pro's cached token used to cost 1.3x flash's and now costs 3.1x it. If you +were on pro mainly because replayed context was almost free there, that +particular reason just got much weaker.

    +

    The cache is still the biggest lever, just a shorter one: a cached input +token now costs about 1/30th of a miss, where it was 1/50th +on flash and 1/120th on pro. Structuring a prompt so the stable part comes +first is worth up to 30x; moving the same work off-peak is worth 2x. Do both, +in that order.

    + +

    We checked the bill, not just the page

    +

    A docs page can update before a billing system does, so this is measured. +On 2026-08-17, in an off-peak window, one deepseek-v4-pro call +with 188,542 cache-miss input tokens settled against a real +RMB account at 0.84 CNY – that is 4.46 CNY per 1M +against the published off-peak 4.5, where the old flat card would have made +it 3.0. Settlement lags the call by a minute or two and arrives in steps, so +poll the balance rather than reading it once.

    +

    One more detail for anyone reconciling the two cards: the USD +figures are the RMB ones converted at a single rate of 6.818 +(it was 6.897 before), with flash's cache-hit cell rounded down – +¥0.05 at 6.818 is $0.00733, published as $0.007.

    + +

    What it changes here: nothing you have to do

    +

    The switch was encoded when it was announced and +gated on its instant, so estimates started using the new card by themselves. +ds pricing names the period you are in right now and when it +next changes; ds models prints the rate in force beside the +model list; the ledger keeps token counts +rather than dollars, so calls made either side of the flip reprice under the +card that was actually real. The embedded docs corpus has been refreshed, so +ds docs show quick_start/pricing is the new page offline.

    +

    The one habit worth adopting: stop quoting a DeepSeek price +without naming the period. There is no longer a single number to +quote, and the difference is a factor of two.

    + +

    2026-08-13 · the price rise has its date and its numberstook effect 2026-08-16

    The other shoe drops. Alongside the V4-Pro GA release, DeepSeek's Models & Pricing page now carries the repricing that the August 6 notice @@ -1745,9 +1872,16 @@ def jstr(s): enough that you should never change history … even more so with deepseek because their cache hit pricing is so low (no stripping old thinking tokens, no rewriting tool results to save context) -(p1necone). The 138x +(p1necone). The cache-hit lever on the cost page is only real if the prefix actually stays put.

    +

    Read this with the date in mind. These +reports were written against the flat card, when pro's cache-read price was +138x below GPT-5.6 Sol's. The 2026-08-16 +repricing raised exactly that number – 6x off-peak, 12x at peak +– leaving the edge at 23x/11x. The advice holds: keep the prefix +stable, never rewrite history. The cost simulations that concluded pro comes +out cheaper do not automatically hold, and are worth re-running.

    Against the other open models

    On GLM-5.2, one report from OpenCode work on an old Unix-clone codebase: diff --git a/site/cost/index.html b/site/cost/index.html index 6ee7095..3e76dee 100644 --- a/site/cost/index.html +++ b/site/cost/index.html @@ -87,20 +87,23 @@

    Cost

    DeepSeek's headline feature is a disk-backed context cache that -makes a repeated prompt prefix roughly fifty times cheaper. That saving is -invisible unless something is counting – so this counts.

    +makes a repeated prompt prefix roughly thirty times cheaper. Since +2026-08-16 a second lever sits on top of it: the hour of the day, worth 2×. +Neither is visible unless something is counting – so this counts.

    The rate card

    -

    USD per 1M tokens, as published on 2026-08-02 and in force before -2026-08-16 16:00 UTC – from that instant DeepSeek bills peak/off-peak -on a new card, and the pricing page carries -the dated schedule:

    +

    USD per 1M tokens. Peak hours are 01:00–04:00 and 06:00–10:00 +UTC daily; every other hour is off-peak at half the peak rate. The +pricing page carries the full schedule and +names the period you are in right now:

    - + - - + + + +
    ModelInput (cached)Input (miss)Output
    ModelPeriodInput (cached)Input (miss)Output
    deepseek-v4-flash$0.0028$0.14$0.28
    deepseek-v4-pro$0.003625$0.435$0.87
    deepseek-v4-flashoff-peak$0.007$0.22$0.66
    peak$0.014$0.44$1.32
    deepseek-v4-prooff-peak$0.022$0.66$1.98
    peak$0.044$1.32$3.96
    @@ -109,22 +112,32 @@

    The rate card

    prints the full schedule with the period you are in right now.

    What the cache is worth

    -

    On flash, a cache hit costs 1/50th of a miss. The same -3,200-token prompt, sent twice:

    +

    On flash, a cache hit costs 1/31st of a miss per token. +The same 2,900-token prompt, sent twice:

    -
    measured, not illustrative
    +
    measured 2026-08-17, off-peak
    $ ds chat "..." --system @prefix.txt
 one
-· flash · 3.2k in · 1 out · ~$0.000450 · 0.84s
+· flash · 2.9k in · 1 out · ~$0.000642 · 580ms
 
 $ ds chat "..." --system @prefix.txt
 two
-· flash · 3.2k in (100% cached) · 1 out · ~$0.000011 · 1.04s
    +· flash · 2.9k in (97% cached) · 1 out · ~$0.000043 · 630ms +
    +

    A 15× drop on the whole call, for changing nothing but sending the +same prefix again. It is 15× and not 31× because the 3% that +changed – the tail of the prompt, and the question itself – +still bills at the miss rate: the realised saving is always the rate ratio +times how much of the prompt you kept identical. Hence the practical rule: +put the stable part of a prompt first – same system +prompt, same files, in the same order – and let the variable part come +last.

    +
    +a trap worth knowing +

    ds tokens counts by sending the text to the FIM endpoint, so +it warms the cache: measure the cache with two chat calls, or +the first one is already a hit.

    -

    A 40× drop, for changing nothing but sending the same prefix again. -The practical rule: put the stable part of a prompt first -– same system prompt, same files, in the same order – and let the -variable part come last.

    The thinking surcharge

    Thinking mode is on by default and adds a fixed template to your input @@ -204,18 +217,16 @@

    What these numbers are not

  • Estimates, not invoices. Computed from the published USD rate card. Your account may bill in another currency – ds balance shows which.
    • -
  • The repricing is dated, and the switch is encoded. -From 2026-08-16 16:00 UTC DeepSeek bills peak/off-peak on a new, higher -card (peak hours 01:00–04:00 and 06:00–10:00 UTC at twice the -off-peak rate). Estimates use the flat card above until that instant and -switch automatically on it – never before, because applying a price -before its effective date would be inventing data. The -pricing page and ds pricing -carry the schedule and the new numbers.
    • -
  • The cache discount narrows at the flip. On the card -above a cached input token costs 1/50th of a miss on flash and 1/120th -on pro; on the card of 2026-08-16 both settle at about 1/30th. Still -the biggest lever on a bill, just a smaller one.
    • +
  • The hour is part of the price. Since 2026-08-16 +16:00 UTC a call costs twice as much inside a peak window as outside it, +so an estimate that ignores the clock is wrong for seven hours a day. +Every call is priced at the card in force at the moment it was made, and +the ledger keeps token counts rather than +dollars, so history reprices correctly under whichever card was real.
    • +
  • The cache discount narrowed at the flip. On the old +flat card a cached input token cost 1/50th of a miss on flash and 1/120th +on pro; on the card above both settle at about 1/30th. Still the biggest +lever on a bill, just a smaller one.
  • Local only. The ledger records calls made by this CLI on this machine. It knows nothing about your other clients.
    • diff --git a/site/news/index.html b/site/news/index.html index 4c238ed..f4ecd58 100644 --- a/site/news/index.html +++ b/site/news/index.html @@ -3,25 +3,25 @@ -DeepSeek API news: the dated repricing, dsh ships, V4-Pro GA - - +DeepSeek API news: the repricing is live, dsh ships, V4-Pro GA + + - - + + - - + + @@ -30,7 +30,7 @@