web_search on the free tier, a dashboard that charts days, first-party upstream health

[lroolle]

Stacked on #4 (feat/status-dashboard) so the diff is only this work; it will retarget to main when #4 merges.

Three changes that each began as a measurement, plus the site copy that follows from them.

web_search works on the free tier

The old refusal argued a server-side tool "performs billed work that never appears in the usage object". Measured against the live API on 2026-08-07: one search request made eleven server-side calls and billed 40,260 input tokens (32k cached) — and the account balance moved by nothing beyond those tokens. Eleven searches at a frontier vendor's $10/1k rate would have been 0.79 CNY against a 14.26 CNY balance, i.e. unmistakable. There is no per-search fee; the whole cost arrives as input tokens the meter already reads exactly.

What the measurement did break is the reservation. meter.Estimate bounded input at one token per body byte — true for every request except this one, because DeepSeek chooses how many pages to read. So:

• a search reserves a 256k input allowance (~6x the observed case);
• it spends one of three daily searches, a new quota dimension, because one search costs about what ten ordinary turns cost and the request counter alone would let one caller take a quarter of the day's budget while looking ordinary;
• every other server-side tool stays refused — unknown work at an unknown price, spent from donated credit.

Within the allowance the budget is still a hard ceiling; the overshoot past it is bounded and stated plainly in DESIGN.md. Reasoning and expiry condition in TASTE.md.

Verified in production (v0.3.0-13): the previously-failing command returns a sourced answer for $0.0013, the 4th search is refused with a message that says ordinary requests still work, and an ordinary request then succeeds.

The dashboard charts a day per bar

It charted output tokens/sec over five minutes — the wrong instrument for a pool serving a few requests an hour, so it was flat at zero whenever anyone looked. The 30-day series comes from quota.History, which reads the same journals the money settles from, so the chart cannot disagree with the ledger. Finished days are immutable and memoised; quiet days are zeroes, not gaps. The live rates stay on the page, demoted to where their usual zero reads as honest rather than broken.

"Is it us or DeepSeek?"

Every upstream round trip is recorded; /v1/status carries the last success, the consecutive-fault streak and the last fault shape. Only DeepSeek's own failures (429/5xx/transport) count as faults — a caller's 4xx would otherwise read as an outage. Before the first forwarded request the state is unknown, not a cheerful all-clear it has not earned.

Deliberately not a scrape of status.deepseek.com. Probed from the production host: no JSON endpoint exists, and the host resolves to an Aliyun Beijing load balancer that accepts TCP and never completes TLS from outside China, while api.deepseek.com answers fine from the same box. A status page also cannot see a route broken only from here. The incident page stays linked for humans.

Site

/economics now cites measured cost per task — Artificial Analysis ($0.03 for v4-flash vs $2.34 Opus 5 and $3.14 Fable 5; $72 vs $3,836 vs $5,455 to run the whole index) and SWE-rebench ($0.15/problem at 40.2% for v4-pro vs $4.40 at 64.5% for Fable 5) — instead of only rate-card arithmetic. It also publishes the two figures that cut against the argument: GPT-5.6 Luna matches flash's intelligence score at $0.05 vs $0.03, and beats v4-pro on both cost and resolve rate on SWE-rebench. The advantage is real against the flagship tier and largely evaporates against the cheap frontier tier; a page that hid that would be propaganda.

The playground gains a web-search toggle (Responses only, the one format DeepSeek offers it on) and deepseek free status shows the ration.

Verify

cd gateway && go test -race ./...          # race-clean
python3 site/build.py --check              # generator and generated HTML agree
curl -s https://freeseek.1lm.io/v1/status | jq '.upstream, .daily_limits_per_user, (.history|length)'

🤖 Generated with Claude Code