the repricing is live — apply it everywhere, and check it against a bill

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
 

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"

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:
 

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 |
 |---|---|---|

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)
 	}
 }

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)
 	}
 }

internal/deepseek/pricing.go

@@ -37,18 +37,27 @@ 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.
 // Source: https://api-docs.deepseek.com/quick_start/pricing (2026-08-13).
 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},

site/bench/index.html

@@ -30,7 +30,7 @@
 <link href="https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@300;400;700&display=swap" rel="stylesheet">
 <script>try{var t=localStorage.getItem('theme');if(t==='dark'||t==='light')document.documentElement.dataset.theme=t}catch(e){}</script>
 <script type="application/ld+json">
-{"@context":"https://schema.org","@type":"FAQPage","mainEntity":[{"@type":"Question","name":"How does DeepSeek-V4-Pro score against other models on agent benchmarks?","acceptedAnswer":{"@type":"Answer","text":"On DeepSeek's own launch-day chart (2026-08-12), V4-Pro-0813 scores Terminal-Bench 2.1 87.9, DeepSWE 62.7, Toolathlon-Verified 74.1, CyberGym 83.3, HLE-with-tools 60.0 and AutomationBench 31.8. It sits in the same cluster as Kimi K3, Claude Fable 5 and Opus 4.8: within a point of Fable 5 on Terminal-Bench (88.0) and CyberGym (83.1), ahead of Opus 4.8 on several execution suites, but behind Kimi K3 on Terminal-Bench, DeepSWE, Toolathlon and DSBench-Hard. These are vendor numbers from one harness and are not yet independently reproduced."}},{"@type":"Question","name":"What did the V4-Pro GA (0813) checkpoint change over the preview?","acceptedAnswer":{"@type":"Answer","text":"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."}},{"@type":"Question","name":"What is the DeepSeek kill line (斩杀线)?","acceptedAnswer":{"@type":"Answer","text":"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."}},{"@type":"Question","name":"Is DeepSeek-V4-Pro better than Claude or GPT for coding agents?","acceptedAnswer":{"@type":"Answer","text":"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."}}]}
+{"@context":"https://schema.org","@type":"FAQPage","mainEntity":[{"@type":"Question","name":"How does DeepSeek-V4-Pro score against other models on agent benchmarks?","acceptedAnswer":{"@type":"Answer","text":"On DeepSeek's own launch-day chart (2026-08-12), V4-Pro-0813 scores Terminal-Bench 2.1 87.9, DeepSWE 62.7, Toolathlon-Verified 74.1, CyberGym 83.3, HLE-with-tools 60.0 and AutomationBench 31.8. It sits in the same cluster as Kimi K3, Claude Fable 5 and Opus 4.8: within a point of Fable 5 on Terminal-Bench (88.0) and CyberGym (83.1), ahead of Opus 4.8 on several execution suites, but behind Kimi K3 on Terminal-Bench, DeepSWE, Toolathlon and DSBench-Hard. These are vendor numbers from one harness and are not yet independently reproduced."}},{"@type":"Question","name":"What did the V4-Pro GA (0813) checkpoint change over the preview?","acceptedAnswer":{"@type":"Answer","text":"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."}},{"@type":"Question","name":"What is the DeepSeek kill line (斩杀线)?","acceptedAnswer":{"@type":"Answer","text":"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."}},{"@type":"Question","name":"Is DeepSeek-V4-Pro better than Claude or GPT for coding agents?","acceptedAnswer":{"@type":"Answer","text":"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."}}]}
 </script>
 <script type="application/ld+json">
 {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"deepseek-cli","item":"https://thevibeworks.github.io/deepseek-cli/"},{"@type":"ListItem","position":2,"name":"bench","item":"https://thevibeworks.github.io/deepseek-cli/bench/"}]}
@@ -178,20 +178,27 @@ <h2 id="kill-line">The kill line (<span lang="zh">斩杀线</span>)</h2>
 <table>
 <thead><tr><th>Per 1M tokens</th><th class="num">v4-flash</th><th class="num">v4-pro</th><th class="num">GPT-5.6 Sol</th><th class="num">pro is cheaper by</th></tr></thead>
 <tbody>
-<tr><td>input, cache miss</td><td class="num">$0.14</td><td class="num">$0.435</td><td class="num">$5.00</td><td class="num">~11.5x</td></tr>
-<tr><td>input, cache hit</td><td class="num">$0.0028</td><td class="num">$0.003625</td><td class="num">$0.50</td><td class="num">~138x</td></tr>
-<tr><td>output</td><td class="num">$0.28</td><td class="num">$0.87</td><td class="num">$30.00</td><td class="num">~34.5x</td></tr>
+<tr><td>input, cache miss</td><td class="num">$0.22 / $0.44</td><td class="num">$0.66 / $1.32</td><td class="num">$5.00</td><td class="num">7.6x / 3.8x</td></tr>
+<tr><td>input, cache hit</td><td class="num">$0.007 / $0.014</td><td class="num">$0.022 / $0.044</td><td class="num">$0.50</td><td class="num">23x / 11x</td></tr>
+<tr><td>output</td><td class="num">$0.66 / $1.32</td><td class="num">$1.98 / $3.96</td><td class="num">$30.00</td><td class="num">15x / 7.6x</td></tr>
 </tbody>
 </table>
 </div>
-<p class="small">DeepSeek prices are the published USD rate card of
-2026-08-02, unchanged at GA; they are a conversion of the RMB card
-(&yen;3 / &yen;0.025 / &yen;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 &ndash; the
-<a href="../pricing/">pricing page</a> 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.</p>
+<p class="small">DeepSeek cells read <strong>off-peak / peak</strong>, on the
+card in force since 2026-08-16 16:00 UTC; they are a conversion of the RMB
+card (pro: &yen;4.5 / &yen;0.15 / &yen;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 <a href="../pricing/">pricing page</a> has the full schedule.</p>
+<div class="note warn">
+<span class="tag">the repricing moved this line</span>
+<p>These ratios were <strong>11.5x / 138x / 34.5x</strong> on the flat card
+that ran until 2026-08-16. The cache-hit column is where the argument
+lived &ndash; a 138x edge is what made replayed context, parallel reviewers
+and long tool loops nearly free &ndash; 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.</p>
+</div>
 <p>The sober version matters as much as the slogan. The kill line is real for
 the <em>middle</em> 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
@@ -212,11 +219,16 @@ <h2 id="practice">What it means in practice</h2>
 <code>ds chat -m deepseek-v4-pro</code> and the
 <a href="../formats/">Anthropic remap</a> make the switch one flag.</li>
 <li><strong>Structure prompts for the cache.</strong> 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; <code>ds usage</code> 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.</li>
+can see whether it is working. This is where the cache-hit column turns from
+a table cell into a bill &ndash; and since the repricing hit that column
+hardest, it is worth more attention now, not less.</li>
+<li><strong>Schedule what can be scheduled.</strong> The same call costs
+half as much outside 01:00&ndash;04:00 and 06:00&ndash;10:00 UTC. Batch
+evaluation, bulk review and overnight agent runs are exactly the workloads
+that can move; <code>ds pricing</code> says which period you are in.</li>
 <li><strong>Measure successful-task cost, not per-token price.</strong> A
 cheaper model that retries five times can cost more than a dearer one that lands
 first. The <a href="../cost/#ledger">ledger</a> stores exact token counts