feat(pricing): V4.1 Flash is deepseek-flash, on a cheaper card

AGENTS.md

@@ -66,7 +66,7 @@ Key flags:
 
 | Flag | Effect |
 | --- | --- |
-| `-m, --model` | `deepseek-v4-flash` (default) or `deepseek-v4-pro` |
+| `-m, --model` | `deepseek-flash` (default) or `deepseek-v4-pro` |
 | `--think on\|off` | thinking mode; default is the API's own, which is on |
 | `-e, --effort low\|high\|max` | reasoning effort |
 | `--max-tokens N` | cap generated tokens |
@@ -106,7 +106,10 @@ deepseek pricing --json          # the schedule and the billing period right now
 ```
 
 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
+01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, bill at twice the
+off-peak rate. Since 2026-09-10 flash is `deepseek-flash` (V4.1) on a
+lower card; the retired names `deepseek-v4-flash` and
+`deepseek-v4-flash-vision-exp` still work and bill as flash. 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.
@@ -163,8 +166,7 @@ deepseek free off            # forget the enrolment on this machine
 
 Free-tier limits, per UTC day: 30 requests, 60K input tokens, 20K output
 tokens, 3 web searches, 4K output per call, 128KB per request body,
-`deepseek-v4-flash` only. A request for pro is **refused, not
-downgraded**. `models` and `status` cost no quota; everything that can
+flash only. A request for pro is **refused, not downgraded**. `models` and `status` cost no quota; everything that can
 generate a token does.
 
 `respond --web-search` works on the free tier and spends one of the three
@@ -265,21 +267,24 @@ 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 about 30× less than uncached ones. Put the
-  stable part of a prompt first — the same system prompt and files across
+- Cached input tokens cost 50× less than uncached ones on flash, 30× on
+  pro. 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
 
-- **Text only.** Images, documents and search-result blocks are rejected
-  by the API in every format.
+- **Text only, from this CLI.** Since 2026-09-10 `deepseek-flash` takes
+  images through chat and Responses, and pro does not; the CLI has no
+  flag that sends one.
 - **The thinking surcharge depends on effort**, and not as documented.
   The template added to your input is fixed per level, constant across
-  prompt length, measured live on 2026-08-05:
+  prompt length, measured live on 2026-08-05 on V4 Flash (retired
+  2026-09-10; `deepseek-flash` is not re-measured, and one probe read +26
+  at `high`):
 
-  | `--effort` | flash | pro |
+  | `--effort` | V4 flash | pro |
   | --- | --- | --- |
   | `none` | +0, thinking off | +0, thinking off |
   | `minimal`, `low` | +0 | +0 |

Makefile

@@ -87,7 +87,8 @@ gateway-check: gateway-test
 # cost estimates, gateway/internal/meter/meter.go for the gateway's
 # budget — because they are separate modules and neither can import the
 # other. Drift would mean the gateway believed it had spent a different
-# amount than it had. This is what catches that.
+# amount than it had. This is what catches that, for the numbers and for
+# the dated instants that switch between them.
 .PHONY: price-check
 price-check:
 	@pattern='CacheHitInput: [0-9.]+, CacheMissInput: [0-9.]+, Output: [0-9.]+'; \
@@ -100,7 +101,15 @@ price-check:
 	   echo "FAIL: the CLI and gateway rate cards have drifted"; \
 	   echo "  cli:     $$cli"; echo "  gateway: $$gw"; exit 1; \
 	 fi; \
-	 echo "rate cards match:"; echo "$$cli" | sed 's/^/  /'
+	 echo "rate cards match:"; echo "$$cli" | sed 's/^/  /'; \
+	 when='time\.Date\([^)]*\)'; \
+	 cli=$$(grep -oE "$$when" internal/deepseek/pricing.go | sort); \
+	 gw=$$(grep -oE "$$when" gateway/internal/meter/meter.go | sort); \
+	 if [ "$$cli" != "$$gw" ]; then \
+	   echo "FAIL: the CLI and gateway switch cards at different instants"; \
+	   echo "  cli:     $$cli"; echo "  gateway: $$gw"; exit 1; \
+	 fi; \
+	 echo "and switch at the same instants:"; echo "$$cli" | sed 's/^/  /'
 
 # The site is generated HTML plus one page that is an application. The
 # generator has its own --check; these cover the parts it cannot see —

README.md

@@ -14,7 +14,7 @@
   <a href="#development"><img src="https://img.shields.io/badge/tests-302%20%C2%B7%2070%25%20covered-brightgreen" alt="302 tests, 70% statement coverage"></a>
   <br>
   <a href="https://api-docs.deepseek.com"><img src="https://img.shields.io/badge/DeepSeek%20API%20docs-2026--08--05-8a2be2" alt="Implemented against the DeepSeek API docs of 2026-08-05"></a>
-  <a href="https://api-docs.deepseek.com/quick_start/pricing"><img src="https://img.shields.io/badge/models-v4--flash%20%7C%20v4--pro-0ea5e9" alt="Models: deepseek-v4-flash and deepseek-v4-pro"></a>
+  <a href="https://api-docs.deepseek.com/quick_start/pricing"><img src="https://img.shields.io/badge/models-flash%20%7C%20v4--pro-0ea5e9" alt="Models: deepseek-flash and deepseek-v4-pro"></a>
   <a href="https://go.dev"><img src="https://img.shields.io/badge/Go-1.25+-00ADD8?logo=go&logoColor=white" alt="Go"></a>
   <a href="LICENSE"><img src="https://img.shields.io/badge/License-MIT-yellow.svg" alt="MIT License"></a>
 </p>
@@ -54,7 +54,7 @@ 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 **~30× cheaper** than an uncached one, and since 2026-08-16 the
+  token **30–50× 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
@@ -171,7 +171,7 @@ is the key, the balance, the network, a proxy, or one endpoint:
 $ deepseek check
 https://api.deepseek.com
 
-ok    GET /models                  141ms  deepseek-v4-flash, deepseek-v4-pro
+ok    GET /models                  141ms  deepseek-flash, deepseek-v4-pro
 ok    GET /user/balance            126ms  18.48 CNY
 ok    POST /chat/completions       651ms  5 in / 1 out
 ok    POST /anthropic/v1/messages  590ms  5 in / 1 out
@@ -376,8 +376,8 @@ 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 $0.007/M against
-$0.22/M for a miss in an off-peak hour — structuring prompts so the
+That last line is the point. On flash, cached input costs $0.003/M
+against $0.15/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.
 
@@ -470,20 +470,23 @@ rather than hides:
   the same number at every level — and at low effort it is not there at
   all, while the model still reasons:
 
-  | `--effort` | flash | pro |
+  | `--effort` | V4 flash | pro |
   | --- | --- | --- |
   | `none` | +0 (thinking off) | +0 (thinking off) |
   | `minimal`, `low` | **+0** | **+0** |
   | `medium`, `high`, `xhigh` | +79 | +0 |
   | `max` | +92 | +79 |
 
   Measured against the live API on 2026-08-05 at two prompt lengths,
-  twice each. `deepseek tokens -e low` will show you the same thing for
+  twice each, on V4 Flash. That model was retired on 2026-09-10 and
+  `deepseek-flash` has not been re-measured: one probe on 2026-09-18
+  read +26 at `high`. `deepseek tokens -e low` will show you the same thing for
   your own text. Two of those levels — `none` and `minimal` — appear in
   no DeepSeek documentation at all; `none` disables thinking exactly as
   `--think off` does.
-- **Text only.** DeepSeek rejects image, document and search-result
-  content blocks in every format.
+- **Text only, from this CLI.** Since 2026-09-10 `deepseek-flash` takes
+  images through chat and Responses, and pro does not; the CLI has no
+  flag that sends one.
 - **The Responses endpoint takes both models** since V4-Pro's official
   release (2026-08-12); it refused pro before that.
 - **FIM caps output at 4K tokens** and ignores thinking entirely.

gateway/DESIGN.md

@@ -281,25 +281,26 @@ scarce operation and gets the conservative boundary.
 ## Limits, and why these numbers
 
 Worst-case cost per anonymous user per day, at flash off-peak rates
-($0.22/M input, $0.66/M output):
+($0.15/M input, $0.60/M output):
 
 ```
-60,000 input  × $0.22/M = $0.0132
-20,000 output × $0.66/M = $0.0132
+60,000 input  × $0.15/M = $0.0090
+20,000 output × $0.60/M = $0.0120
                           -------
-                          $0.0264 / user / day  at full burn, off-peak
-                          $0.0528 / user / day  at full burn, peak
+                          $0.0210 / user / day  at full burn, off-peak
+                          $0.0420 / user / day  at full burn, peak
 ```
 
-So a $1/day budget serves ~38 users burning *everything* off-peak, ~19 at
+So a $1/day budget serves ~47 users burning *everything* off-peak, ~23 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
+input and a few hundred output tokens, about $0.0005. Roughly **1,900
+ordinary turns per dollar off-peak, 950 at peak.**
+
+These are the numbers of the V4.1 Flash card of 2026-09-10. The card of
+2026-08-16 16:00 UTC ($0.22/$0.66) made the same worst case
+$0.0264/user/day, which had roughly halved the free tier's reach against
+the flat card before it ($0.14/$0.28, $0.014/user/day). The quotas below
+have not been retuned for either — 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.

gateway/internal/meter/meter.go

@@ -55,29 +55,57 @@ var RepriceAt = time.Date(2026, time.August, 16, 16, 0, 0, 0, time.UTC)
 // The drift here runs the other way to the repricing's: over-charging our
 // own budget is safe for the credit pool but it is still wrong, and it
 // makes /economics report a cost-per-task the account never paid. The
-// same date gates the CLI's copy; `make price-check` guards the numbers,
-// and this comment is the only thing guarding the clock.
+// same date gates the CLI's copy; `make price-check` guards the numbers
+// and, since the V4.1 cut added a third instant, the dates too.
 var weekendOffPeakAt = time.Date(2026, time.August, 22, 16, 0, 0, 0, time.UTC)
 
+// v41At is when the Flash card dropped with the DeepSeek-V4.1-Flash
+// release of 2026-09-10; Pro's card did not move. The instant is
+// INFERRED, as in the CLI's copy (internal/deepseek/pricing.go, V41At,
+// which records how): upstream published a date and no time. A cut
+// cannot drain the pool the way the repricing could, but it can make
+// /economics report a cost nobody paid.
+var v41At = time.Date(2026, time.September, 10, 11, 0, 0, 0, time.UTC)
+
 // beijing is the vendor's clock; no daylight saving since 1991.
 var beijing = time.FixedZone("CST", 8*60*60)
 
+// The cards are keyed by tier, not by model name. The API has accepted
+// three flash names (deepseek-v4-flash, deepseek-v4-flash-vision-exp,
+// deepseek-flash) and bills them all as flash, so cardFor maps a name
+// onto its tier and each card says what that tier cost in its era.
+const (
+	flashTier = "flash"
+	proTier   = "pro"
+)
+
 // ratesFlat is the card published 2026-08-02, in force before RepriceAt.
 var ratesFlat = map[string]Price{
-	"deepseek-v4-flash": {CacheHitInput: 0.0028, CacheMissInput: 0.14, Output: 0.28},
-	"deepseek-v4-pro":   {CacheHitInput: 0.003625, CacheMissInput: 0.435, Output: 0.87},
+	flashTier: {CacheHitInput: 0.0028, CacheMissInput: 0.14, Output: 0.28},
+	proTier:   {CacheHitInput: 0.003625, CacheMissInput: 0.435, Output: 0.87},
+}
+
+// ratesV4OffPeak is the base card from RepriceAt until v41At; during
+// peakWindows every billing item costs peakMultiplier times these numbers.
+var ratesV4OffPeak = map[string]Price{
+	flashTier: {CacheHitInput: 0.007, CacheMissInput: 0.22, Output: 0.66},
+	proTier:   {CacheHitInput: 0.022, CacheMissInput: 0.66, Output: 1.98},
+}
+
+// ratesV41OffPeak is the base card from v41At on: Flash cut on every
+// item, Pro unchanged. Same peak rule.
+var ratesV41OffPeak = map[string]Price{
+	flashTier: {CacheHitInput: 0.003, CacheMissInput: 0.15, Output: 0.6},
+	proTier:   {CacheHitInput: 0.022, CacheMissInput: 0.66, Output: 1.98},
 }
 
-// ratesOffPeak is the base card from RepriceAt on; during peakWindows
-// every billing item costs peakMultiplier times these numbers.
-var ratesOffPeak = map[string]Price{
-	"deepseek-v4-flash": {CacheHitInput: 0.007, CacheMissInput: 0.22, Output: 0.66},
-	// Released 2026-08-21, after the switchover, so it has no flat row
-	// above. Priced identically to flash. Carried here even though the
-	// policy allowlist does not admit it, so that a metering gap can never
-	// be the reason it gets served for free.
-	"deepseek-v4-flash-vision-exp": {CacheHitInput: 0.007, CacheMissInput: 0.22, Output: 0.66},
-	"deepseek-v4-pro":              {CacheHitInput: 0.022, CacheMissInput: 0.66, Output: 1.98},
+// offPeakCardAt is the time-of-day base card in force at t, which must
+// not be before RepriceAt.
+func offPeakCardAt(t time.Time) map[string]Price {
+	if t.Before(v41At) {
+		return ratesV4OffPeak
+	}
+	return ratesV41OffPeak
 }
 
 const peakMultiplier = 2.0
@@ -126,24 +154,21 @@ func PriceAt(model string, t time.Time) Price {
 	if t.Before(RepriceAt) {
 		return cardFor(ratesFlat, model)
 	}
-	p := cardFor(ratesOffPeak, model)
+	p := cardFor(offPeakCardAt(t), model)
 	if inPeak(t) {
 		p = scale(p, peakMultiplier)
 	}
 	return p
 }
 
 func cardFor(cards map[string]Price, model string) Price {
-	if p, ok := cards[model]; ok {
-		return p
-	}
 	switch {
 	case strings.Contains(model, "pro") || model == "":
-		return cards["deepseek-v4-pro"]
+		return cards[proTier]
 	case strings.Contains(model, "flash"):
-		return cards["deepseek-v4-flash"]
+		return cards[flashTier]
 	default:
-		return cards["deepseek-v4-pro"]
+		return cards[proTier]
 	}
 }
 
@@ -228,10 +253,21 @@ func ceilingAt(model string, t time.Time) Price {
 	if t.Add(inFlight).Before(RepriceAt) {
 		return cardFor(ratesFlat, model)
 	}
+	// The base card is the admission instant's. The one change since
+	// the repricing, v41At, was a cut, so a request admitted before it is
+	// already reserved at the dearer card; a card that ever goes UP needs
+	// the same look-ahead the flip gets above.
 	if !t.Before(RepriceAt) && !peakTouches(t, inFlight) {
-		return cardFor(ratesOffPeak, model)
+		return cardFor(offPeakCardAt(t), model)
+	}
+	return scale(cardFor(offPeakCardAt(maxTime(t, RepriceAt)), model), peakMultiplier)
+}
+
+func maxTime(a, b time.Time) time.Time {
+	if a.After(b) {
+		return a
 	}
-	return scale(cardFor(ratesOffPeak, model), peakMultiplier)
+	return b
 }
 
 // peakTouches reports whether any instant of [t, t+d] falls in a peak

gateway/internal/meter/meter_test.go

@@ -233,7 +233,7 @@ func TestEstimateCeilingCoversTheNextPeriod(t *testing.T) {
 	// And far from any boundary, off-peak reserves at the off-peak card,
 	// not at a permanent doubling.
 	quiet := time.Date(2026, 8, 17, 12, 0, 0, 0, time.UTC)
-	offPeak := costWith(cardFor(ratesOffPeak, model), Usage{InputTokens: 401, OutputTokens: 1000 + reasoningAllowance})
+	offPeak := costWith(cardFor(ratesV4OffPeak, model), Usage{InputTokens: 401, OutputTokens: 1000 + reasoningAllowance})
 	if est := EstimateAt(model, 400, 1000, false, quiet); math.Abs(est-offPeak) > 1e-12 {
 		t.Errorf("quiet off-peak reservation %v, want the off-peak card's %v", est, offPeak)
 	}

gateway/internal/meter/weekend_test.go

@@ -48,16 +48,42 @@ func TestWeekendIsReadOnTheVendorClock(t *testing.T) {
 	}
 }
 
-func TestVisionVariantIsMetered(t *testing.T) {
+func TestEveryFlashNameIsMeteredAsFlash(t *testing.T) {
 	// A model priced upstream but missing from this card would be served
-	// for free out of the budget pool.
-	at := time.Date(2026, 8, 26, 12, 0, 0, 0, time.UTC)
-	vision := PriceAt("deepseek-v4-flash-vision-exp", at)
-	flash := PriceAt("deepseek-v4-flash", at)
-	if vision != flash {
-		t.Errorf("vision %+v != flash %+v", vision, flash)
-	}
-	if vision.CacheMissInput == 0 {
-		t.Error("vision meters at zero")
+	// for free out of the budget pool, or, the other way, at pro rates
+	// that /economics then reports as spent. The API has billed three
+	// names as flash; each must land on the flash card of its instant.
+	for _, at := range []time.Time{
+		time.Date(2026, 8, 10, 12, 0, 0, 0, time.UTC), // flat
+		time.Date(2026, 8, 26, 12, 0, 0, 0, time.UTC), // V4, off-peak
+		time.Date(2026, 9, 14, 2, 0, 0, 0, time.UTC),  // V4.1, peak
+	} {
+		flash := PriceAt("deepseek-v4-flash", at)
+		for _, name := range []string{"deepseek-flash", "deepseek-v4-flash-vision-exp", "deepseek-v4-flash-free"} {
+			if got := PriceAt(name, at); got != flash {
+				t.Errorf("%s at %v = %+v, want flash's %+v", name, at, got, flash)
+			}
+		}
+		if flash.CacheMissInput == 0 {
+			t.Errorf("flash meters at zero at %v", at)
+		}
+	}
+}
+
+func TestV41CutFlashAndLeftPro(t *testing.T) {
+	before := v41At.Add(-time.Minute) // 10:59 UTC Thursday, off-peak
+	if got := PriceAt("deepseek-flash", before); got != (Price{CacheHitInput: 0.007, CacheMissInput: 0.22, Output: 0.66}) {
+		t.Errorf("flash just before V4.1 = %+v, want the V4 card", got)
+	}
+	if got := PriceAt("deepseek-flash", v41At); got != (Price{CacheHitInput: 0.003, CacheMissInput: 0.15, Output: 0.6}) {
+		t.Errorf("flash at V4.1 = %+v, want the V4.1 card", got)
+	}
+	if a, b := PriceAt("deepseek-v4-pro", before), PriceAt("deepseek-v4-pro", v41At); a != b {
+		t.Errorf("pro moved at V4.1: %+v -> %+v", a, b)
+	}
+	// A reservation made before the cut must still cover a settlement
+	// after it: the ceiling may not be beaten by the clock.
+	if est, after := EstimateAt("deepseek-flash", 400, 1000, false, before), EstimateAt("deepseek-flash", 400, 1000, false, v41At); est < after {
+		t.Errorf("admitted before the cut the reservation %v is under the post-cut %v", est, after)
 	}
 }

internal/cli/anthropic.go

@@ -48,7 +48,7 @@ check how your DeepSeek key behaves for those tools before pointing them
 at it.
 
 Claude model names are accepted and remapped server-side: claude-opus* to
-deepseek-v4-pro, claude-sonnet*/claude-haiku* to deepseek-v4-flash, and
+deepseek-v4-pro, claude-sonnet*/claude-haiku* to deepseek-flash, and
 anything unrecognised to flash. The usage line shows both names so the
 cost is traceable to the model that actually ran.
 

internal/cli/chat.go

@@ -64,7 +64,7 @@ under reasoning_content.
 	}
 
 	fl := cmd.Flags()
-	fl.StringVarP(&f.model, "model", "m", deepseek.ModelFlash, "model: deepseek-v4-flash or deepseek-v4-pro")
+	fl.StringVarP(&f.model, "model", "m", deepseek.ModelFlash, "model: deepseek-flash or deepseek-v4-pro")
 	fl.StringVarP(&f.system, "system", "s", "", "system prompt, or @file")
 	fl.StringVar(&f.think, "think", "", "thinking mode: on or off (default: the API's own default, on)")
 	fl.StringVarP(&f.effort, "effort", "e", "", "reasoning effort: low, high, max (pro promotes low to high)")