feat: deepseek-flash everywhere, and web_search is gone

AGENTS.md

@@ -88,13 +88,18 @@ deepseek fim "def f():" --suffix "    return x"
 Reach for `anthropic` or `respond` when the task is specifically about
 those wire formats. For a plain answer, use `chat`.
 
-Only `respond` has JSON Schema output and server-side web search:
+Only `respond` has JSON Schema output:
 
 ```bash
 deepseek respond "Berlin" -s "Return city and country." --schema @city.json
-deepseek respond "what shipped in Go 1.26" --web-search
 ```
 
+There is no server-side web search any more. DeepSeek removed the
+Responses API's `web_search` tool on 2026-09-10 and now ignores it, so
+`respond --web-search` exits 1 with that reason and sends nothing. To
+ground an answer, search yourself and pass the results with `--file` or
+stdin.
+
 ### Account and cost
 
 ```bash
@@ -165,15 +170,12 @@ 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,
-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
-daily searches. It is rationed that tightly because DeepSeek reads whole
-pages into the prompt — one measured search request billed 40K input
-tokens, about ten ordinary turns — so treat it as a few lookups a day, not
-a research loop. Other server-side tools are still refused.
+tokens, 4K output per call, 128KB per request body, `deepseek-flash`
+only (the retired names `deepseek-v4-flash` and
+`deepseek-v4-flash-vision-exp` are the same model). A request for pro is
+**refused, not downgraded**. `models` and `status` cost no quota;
+everything that can generate a token does. Server-side tools are refused,
+`web_search` with the reason that DeepSeek removed it.
 
 Errors from the gateway carry `"type":"free_tier_*"` and a message that
 already contains the next step — do not append DeepSeek's own advice to

README.md

@@ -134,7 +134,7 @@ The free tier relays your prompts to DeepSeek through a gateway run by
 this project. No account, no API key.
 
   gateway   https://freeseek.1lm.io
-  model     deepseek-v4-flash
+  model     deepseek-flash
   per day   30 requests · 60k input · 20k output tokens
   privacy   prompts and completions are relayed to DeepSeek and are not
             stored or logged by this gateway; only token counts and cost
@@ -189,7 +189,7 @@ One command per endpoint, named for what it does.
 | --- | --- | --- |
 | `chat` | `POST /chat/completions` | The default. OpenAI format, the one most tools speak. |
 | `anthropic` | `POST /anthropic/v1/messages` | The format Claude Code and the Anthropic SDKs speak. |
-| `respond` | `POST /responses` | The format Codex speaks. JSON Schema output and server-side `web_search` live only here. |
+| `respond` | `POST /responses` | The format Codex speaks. JSON Schema output lives only here. |
 | `fim` | `POST /beta/completions` | Fill in the middle — the shape editors use for inline completion. |
 | `models` | `GET /models` | Available models, joined with the published rate card. |
 | `balance` | `GET /user/balance` | What is left, per currency. |
@@ -209,7 +209,7 @@ One command per endpoint, named for what it does.
 deepseek chat "why is the sky blue"
 git diff | deepseek chat "write a commit message"
 deepseek chat "explain" --file server.go --file server_test.go
-deepseek chat "review this" --model deepseek-v4-pro --effort max
+deepseek chat "review this" --effort max
 deepseek chat "summarise" --system @house-style.md
 ```
 
@@ -322,11 +322,11 @@ page keeps the upstream URL it was converted from.
 ```bash
 # Anthropic Messages. Claude model names are accepted and remapped
 # server-side; the usage line shows both so cost stays traceable.
-deepseek anthropic "hello" --model claude-opus-4-1
-# · claude-opus-4-1→pro · 10 in · 8 out · ~$0.000011 · 0.9s
+deepseek anthropic "hello" --model claude-sonnet-4-5
+# · claude-sonnet-4-5→flash · 10 in · 8 out · ~$0.000006 · 0.9s
 
-# Responses: JSON Schema output, and a web_search tool DeepSeek runs
-deepseek respond "what shipped in Go 1.26" --web-search
+# Responses: JSON Schema output. Its server-side web_search tool was
+# removed upstream on 2026-09-10; --web-search now exits with that reason.
 deepseek respond "Berlin" -s "Return city and country." --schema @city.json
 
 # FIM: prefix in, suffix optional, the middle comes back
@@ -366,13 +366,13 @@ JSONL ledger:
 
 ```console
 $ deepseek usage --since 7d
-                   CALLS  IN     CACHED  OUT    COST
-deepseek-v4-flash  184    2.1M   78%     94k    $0.19
-deepseek-v4-pro    12     88k    41%     11k    $0.03
-total              196    2.2M   77%     105k   $0.22
+                 CALLS  IN     CACHED  OUT    COST
+deepseek-flash   184    2.1M   78%     94k    $0.13
+deepseek-v4-pro  12     88k    41%     11k    $0.06
+total            196    2.2M   77%     105k   $0.19
 
 by format: chat 170, anthropic 14, responses 8, fim 4
-context cache saved ~$0.23 (1.7M of 2.2M prompt tokens replayed)
+context cache saved ~$0.26 (1.7M of 2.2M prompt tokens replayed)
 costs are estimates from the published USD rate card, not billed amounts
 ```
 

TASTE.md

@@ -75,8 +75,9 @@ return genuinely different objects — `choices[].message` versus
 with opposite cache conventions. A unified command has two ways out and
 both are bad: normalize, and `--json` lies about what the API sent; or
 leak, and one command emits three shapes depending on a flag. Meanwhile
-format-specific parameters (`web_search` and JSON Schema only exist on
-Responses, `prefix` only on beta chat) have to hide behind conditional
+format-specific parameters (`web_search` — until DeepSeek removed it on
+2026-09-10 — and JSON Schema only exist on Responses, `prefix` only on
+beta chat) have to hide behind conditional
 validation. One verb, four behaviours, and the user still has to know
 which format they are in — the surface got smaller while the task got
 harder.
@@ -472,3 +473,23 @@ the ration are both sized to a single measurement and should be re-measured
 when the tool changes. If a search request is ever observed above 256k
 input tokens in production, that is the signal to raise the allowance
 rather than to quietly accept the overshoot.
+
+**Expired 2026-09-18, by upstream.** DeepSeek removed `web_search` from
+the Responses API on 2026-09-10, with V4.1 Flash: the guide now lists it
+among the built-in tools that are *ignored*, and its streaming events are
+gone. A probe on 2026-09-18 confirmed it — the tool was accepted and
+echoed, no search ran, and the request billed 21 input tokens where the
+measured search above billed 40,260. The ration and the allowance
+therefore priced searches that could no longer happen, and the flag sent
+a request that answered from the model's memory to a caller who believed
+it had searched.
+
+What replaced it: the gateway refuses the tool with a 400 that says why,
+`respond --web-search` exits 1 with the same reason and sends nothing,
+and `DailySearches` and `searchInputAllowance` are deleted. The flag is
+kept, hidden, so a script that still passes it learns the reason instead
+of reading "unknown flag" as a bug here. Refusing rather than stripping
+the tool and forwarding is the same judgement as refusing pro rather than
+downgrading it: a caller who asked for something must not be handed less
+without being told. If DeepSeek brings a server-side search back, this
+entry and the one above are the arithmetic to redo, not to restore.

gateway/DESIGN.md

@@ -310,7 +310,6 @@ expensive seven hours.
 | `DSGATE_ANON_DAILY_REQUESTS` | 30 | enough to be useful for a day's work, not enough to script against |
 | `DSGATE_ANON_DAILY_INPUT_TOKENS` | 60000 | ~150 pages of context per day |
 | `DSGATE_ANON_DAILY_OUTPUT_TOKENS` | 20000 | the expensive side; the real cap |
-| `DSGATE_ANON_DAILY_SEARCHES` | 3 | a `web_search` request costs ~10 ordinary turns; the request counter alone would let one caller take a quarter of the day |
 | `DSGATE_ANON_MAX_TOKENS` | 4096 | clamps a single response, bounding overshoot |
 | `DSGATE_MAX_BODY_BYTES` | 131072 | ~32K tokens; bounds the input side of overshoot |
 | `DSGATE_DAILY_BUDGET_USD` | 1.00 | the circuit breaker; the number that actually protects us |
@@ -323,22 +322,29 @@ expensive seven hours.
 | `DSGATE_TOKEN_TTL_DAYS` | 7 | identities age out instead of accumulating |
 | `DSGATE_BALANCE_CHECK_MINUTES` | 15 | the ledger's "we have credit" is checked against the real account |
 
-**`web_search` is carried, and rationed.** Measured on 2026-08-07, one
-search request made eleven server-side calls and billed 40,260 input
-tokens with no separate per-search fee — so its whole cost arrives as
-input tokens the meter already reads. What it breaks is the *reservation*,
-which bounded input at one token per body byte: DeepSeek chooses how many
-pages to read, so a search's input is upstream-controlled. Hence a 256k
-input allowance at admission (about 6x the observed case) plus the daily
-ration above. Within that allowance the budget is still a hard ceiling;
-past it a search can overshoot by the difference, bounded by how many
-distinct callers can be mid-search at once. Every other server-side tool
-stays refused: unknown work at an unknown price, spent from donated
-credit. Reasoning and the expiry condition are in `TASTE.md`.
-
-**Free tier is flash only.** Pro is 3x the price and the request is
-*rejected*, not silently downgraded — a user who asked for pro and got
-flash without being told would draw wrong conclusions and blame the model.
+**No server-side tool is carried.** From 2026-08-07 to 2026-09-18 the
+free tier carried `web_search`, with a 3/day ration and a 256k input
+allowance at admission, because one measured search billed 40,260 input
+tokens. DeepSeek removed the tool from the Responses API on 2026-09-10,
+with V4.1 Flash: the guide now lists `web_search` among the built-in
+tools that are *ignored*, and a probe on 2026-09-18 billed 21 input
+tokens where a search used to bill ~40K. Forwarding it would bill an
+answer from the model's memory to a caller who believes it searched, so
+it is refused with that reason, and the ration and the allowance are
+gone. Every other server-side tool stays refused: unknown work at an
+unknown price, spent from donated credit. The history is in `TASTE.md`.
+
+**Free tier is `deepseek-flash` only.** It is DeepSeek's current Flash
+model (V4.1, 2026-09-10); `deepseek-v4-flash` and
+`deepseek-v4-flash-vision-exp` are retired names upstream serves as the
+same model, so the gateway accepts them as that model and pins the
+request to the name upstream lists. A `DSGATE_MODEL` that still says
+`deepseek-v4-flash` is served as `deepseek-flash` for the same reason —
+otherwise `/models`, which upstream now answers with the new name only,
+would filter to nothing. Pro is 3.3x flash's price per output token and
+the request is *rejected*, not silently downgraded — a user who asked for
+pro and got flash without being told would draw wrong conclusions about
+the model.
 
 ---
 
@@ -408,7 +414,7 @@ stylesheet; the donation path is a private message to a human.
 
 The key pool answers "what if our key runs out". It does not answer the
 prior question — why is a chat request costing us anything at all, when
-OpenCode Zen serves `deepseek-v4-flash-free` for nothing.
+OpenCode Zen served `deepseek-v4-flash-free` for nothing.
 
 So there are two upstreams now, tried in order. The indirection is not
 "a list of interchangeable backends"; a second lane earns its keep only
@@ -419,7 +425,7 @@ the model by, what it costs, and how often it says no.
 Measured against Zen on 2026-08-12:
 
 - `/chat/completions` works, reporting usage in both streamed and
-  buffered form. `/responses` answers but rejects a server-side
+  buffered form. `/responses` answers but rejected a server-side
   `web_search` tool. `/anthropic/v1/messages`, `/beta/completions` and
   `/user/balance` are 404.
 - About one sequential request in five comes back
@@ -428,11 +434,23 @@ Measured against Zen on 2026-08-12:
 - The model is `deepseek-v4-flash-free` there. Our callers never learn
   that: `policy.Retarget` renames the field on the way out, and the
   allowlist, the `/models` list and the client contract all keep saying
-  `deepseek-v4-flash`.
+  the served name — `deepseek-flash` since 2026-09-18.
 - Zen's own docs say free-lane data **may be used to improve the model**.
   That is a different promise from the paid path's and belongs in the
   user-facing copy, not just here.
 
+**Re-measured 2026-09-18: the lane is dead.** Zen still lists
+`deepseek-v4-flash-free` and answers every request with
+`Model is unavailable` (V4 Flash was retired upstream on 2026-09-10).
+The fallback keeps the service correct — every chat still lands on the
+DeepSeek key — but each one pays a refused round trip, and its prompt
+still reaches Zen. So the operator step is to unset `OPENCODE_API_KEY`
+until Zen serves a free V4.1 model. The rename above is why that check
+cannot be skipped: callers are told `deepseek-flash` whatever the lane
+runs, so a revived free lane has to be shown to serve V4.1 (the 53-token
+billed prompt offset between the V4 and V4.1 families is one test)
+before it goes back on.
+
 A 20% refusal rate is why the lane is first-choice rather than the whole
 service, and why the fallback must happen before a single byte reaches
 the client — after `WriteHeader` the status line is spent. Any non-2xx

gateway/README.md

@@ -37,9 +37,10 @@ What it changes about a request, and nothing else:
 
 | | |
 |---|---|
-| `model` | pinned to flash; a pro request is **refused**, not downgraded |
+| `model` | pinned to `deepseek-flash`; the retired flash names are the same model; a pro request is **refused**, not downgraded |
 | `max_tokens` | clamped to the free-tier ceiling |
 | `n`, `best_of` | refused above 1 — they multiply the cost of one admitted request |
+| server-side tools | refused; `web_search` with the reason: DeepSeek removed it on 2026-09-10 and now ignores it |
 | user identity | overwritten with the token's subject |
 
 That last one is not a nicety. DeepSeek documents `user_id` as the
@@ -63,15 +64,25 @@ Measured against Zen on 2026-08-12, which is why the lane is this narrow:
 |---|---|
 | refusal rate | ~20% of sequential requests, `429 FreeUsageLimitError` |
 | `/chat/completions` | works; usage reported streamed and buffered |
-| `/responses` | answers, but rejects a server-side `web_search` tool |
+| `/responses` | answers, but rejected a server-side `web_search` tool |
 | `/anthropic/v1/messages`, `/beta/completions`, `/user/balance` | 404 |
 | the model's name there | `deepseek-v4-flash-free`, aliased at the last moment |
 | privacy | Zen says free-lane data **may be used to improve the model** |
 
 So it carries `chat` and nothing else. FIM, the Anthropic and Responses
-formats, web search and the model list all go straight to DeepSeek, and
-the caller's contract does not change: they ask for `deepseek-v4-flash`,
-by that name, on every route.
+formats and the model list all go straight to DeepSeek, and the caller's
+contract does not change: they ask for `deepseek-flash`, by that name, on
+every route.
+
+**Measured 2026-09-18: the lane carries nothing right now.** Zen still
+lists `deepseek-v4-flash-free`, but answers every request with
+`Model is unavailable` — V4 Flash was retired upstream on 2026-09-10 and
+Zen has no free V4.1 model. Every chat request therefore pays one refused
+round trip and still lands on the DeepSeek key, and its prompt still
+reaches Zen on the way. Leave `OPENCODE_API_KEY` unset until Zen serves a
+free `deepseek-flash`, and check that the model it serves is V4.1 before
+turning it back on: the gateway renames the model on the way out, so
+callers would be told `deepseek-flash` whatever Zen actually runs.
 
 The interesting consequence is what happens when the money runs out.
 A request that the free lane can serve is admitted **past** the daily

gateway/cmd/dsgate/main.go

@@ -76,15 +76,14 @@ behaves exactly as it did before this existed.
   DSGATE_ADDR                listen address                   (:8787)
   DSGATE_STATE_DIR           journal, secret, revocations     (./state)
   DSGATE_SECRET              token signing secret, hex        (generated and persisted)
-  DSGATE_MODEL               the only model served            (deepseek-v4-flash)
+  DSGATE_MODEL               the only model served            (deepseek-flash)
   DSGATE_ANNOUNCE            public URL, shown to clients
 
 Per-user daily limits:
 
   DSGATE_ANON_DAILY_REQUESTS       (30)
   DSGATE_ANON_DAILY_INPUT_TOKENS   (60000)
   DSGATE_ANON_DAILY_OUTPUT_TOKENS  (20000)
-  DSGATE_ANON_DAILY_SEARCHES       (3)      server-side web searches per user
   DSGATE_ANON_MAX_TOKENS           (4096)    per-request output cap
   DSGATE_MAX_BODY_BYTES            (131072)  per-request body cap
   DSGATE_REQUESTS_PER_MINUTE       (20)      per-address burst
@@ -152,7 +151,6 @@ func run() error {
 		DailyRequests:     envInt("DSGATE_ANON_DAILY_REQUESTS", 30),
 		DailyInputTokens:  envInt("DSGATE_ANON_DAILY_INPUT_TOKENS", 60000),
 		DailyOutputTokens: envInt("DSGATE_ANON_DAILY_OUTPUT_TOKENS", 20000),
-		DailySearches:     envInt("DSGATE_ANON_DAILY_SEARCHES", 3),
 		DailyBudgetUSD:    envFloat("DSGATE_DAILY_BUDGET_USD", 1.00),
 		TotalBudgetUSD:    envFloat("DSGATE_TOTAL_BUDGET_USD", 20.00),
 	}
@@ -180,7 +178,7 @@ func run() error {
 		UpstreamBaseURL:          env("DSGATE_UPSTREAM_BASE_URL", "https://api.deepseek.com"),
 		UpstreamKeys:             keys,
 		KeyStatePath:             filepath.Join(stateDir, "donated-keys.json"),
-		Model:                    env("DSGATE_MODEL", "deepseek-v4-flash"),
+		Model:                    env("DSGATE_MODEL", "deepseek-flash"),
 		FreeBaseURL:              env("DSGATE_FREE_BASE_URL", "https://opencode.ai/zen/v1"),
 		FreeKeys:                 freeKeys,
 		FreeModel:                env("DSGATE_FREE_MODEL", "deepseek-v4-flash-free"),

gateway/deploy/README.md

@@ -14,7 +14,7 @@ DSGATE_UPSTREAM_KEY=sk-... dsgate
 ```
 
 That listens on `:8787`, writes state to `./state`, serves
-`deepseek-v4-flash`, and stops spending at **$1.00 a day / $20.00
+`deepseek-flash` (DeepSeek-V4.1-Flash), and stops spending at **$1.00 a day / $20.00
 total**. Point a CLI at it:
 
 ```bash
@@ -39,8 +39,18 @@ anything. `GET /v1/status` reports the share it is carrying under
 Two things to know before turning it on. Zen's free lane says the prompts
 it sees **may be used to improve the model**, which is a different promise
 from the one the paid path makes, so say so wherever you tell users where
-their prompts go. And it only carries `chat`: FIM, the Anthropic and
-Responses formats, and web search still spend real credit.
+their prompts go. And it only carries `chat`: FIM and the Anthropic and
+Responses formats still spend real credit.
+
+**As of 2026-09-18 Zen answers its free DeepSeek model with
+`Model is unavailable` on every request** (V4 Flash was retired upstream
+on 2026-09-10), so the lane saves nothing and adds a refused round trip
+to every chat. Leave `OPENCODE_API_KEY` unset until that changes; see the
+gateway README.
+
+A `DSGATE_MODEL` written before 2026-09-10 (`deepseek-v4-flash`) still
+works: the gateway serves it under the name upstream uses today,
+`deepseek-flash`, and says so in its boot log.
 
 ## Docker