Models & APIs — by function

The console's Models & APIs tab regrouped onto the ledger's eight groups — speech · fp · tool · kit · persona · photo · room · app — in the Usage & cost tab's order and colours, so a figure on Usage ▸ By function and the knob that moves it are one hop apart. Every leaf of the vocabulary gets a row, including the eleven that have no knob today.

1 · Two maps of one thing

A reader of the console holds two different groupings of the same money. Usage stacks spend by the ledger's call kind — the closed group.leaf vocabulary of usage-dashboard §3, eight coloured groups. Models lays the knobs out by which part of the product they sit in — six uncoloured surfaces. So「kit cost $3.10 this week」sends the reader to Models to find… no kit.

The eleven missing leaves — kit.* ×5, persona.* ×4, photo.* ×3 less the eyes' price row, room.translate, speech.interrogate, tool.promised — are not missing by accident: their model is a constant in code, so there was never a knob to draw. But the page is read as an inventory (「every model & search API, by role」 is its own subtitle), and an inventory that omits a third of the ledger is a page that cannot answer「what runs persona.dream?」.

The three rules of the regrouping

1. Same eight, same order, same colours. The groups are U_GROUPS — the usage tab's own list, in the same file. A group rail under the tab title is the usage chart's legend made a navigation: one chip per group, the chart's colour, the leaf count; click to jump, and it follows the scroll.
2. Every leaf has a card, and every card is the same card. The key is the stored kind (room.translate), then the name (Translate on divergence); the rows are the console's radio rows. Where a setting exists the rows are the choices you have today, untouched. Where the model is a constant in code the card has one row, one radio, checked — the fact, in the same format as a choice. Where inherit is the default, the Inherit row is written out, followed by the model it resolves to. Every row marks thinking — thinking off / thinking on as a word, with the switch beside it where it is a choice — so a call that cannot think says so instead of saying nothing. No descriptive text anywhere; the server reports the constants, the page never names a model itself.
3. Prices stay one surface — a ninth, neutral stop at the end. It is not a group; it is what every group above is billed from.

2 · The tab

Markup and classes lifted from lib/admin.html, set in sans; the rail, the group header, the kind-as-key card header and the thinking mark are the new pieces; every row is the console's .mrow.

3 · The map — every leaf, its row

The build follows this table: 28 stored kinds, where each sits today, and what its card holds. choice = an existing setting, moved as-is · one radio = a constant in code, one checked row · inherit = an Inherit row, checked.

4 · What changes (the plan, as built)

• Render order only, on the front. loadModels() emits the groups in U_GROUPS order with the new .fgrp header; every card, every POST endpoint and every change handler is the same code — a card moves, it does not change.
• The one-radio cards come from the server. GET /api/admin/models grows a fixed block — {leaf: {model}}, or {inherit: true} — read from the constants themselves (MEMORY_MODEL, PHOTO_MODEL, PHOTO_DENSE_EYES, ACT_GATE_MODEL, …). The renderer draws it as the same .mrow it draws a choice with — one radio, checked, no handler. The page never hard-codes a model name, so a constant that moves in code moves on the page the same commit.
• The prop master: Off · V4 Pro · V4.1 Flash — one radio group; Off posts prop-gate on=false, a model posts on=true + the model. Same two endpoints as today, one control instead of two.
• One lint, because the vocabulary is closed. The smoketest checks that every kind the ledger can book (db.KIND_ALIAS values + the §3 table) appears as a .kind key in the Models render — a new leaf with no row fails the gate, the way a new kind with no colour already does.
• The rail's colours are the chart's. --s1…8 already live in theme.css for the usage tab; the rail, the group name and the kind key all wear them (a --fc per group section). No new hue anywhere.
• Sans for all. The tab drops mono — card header, kind key, chips, the thinking mark, the price table — everything is var(--sans); the tab title keeps the console's serif. 「reasoning」becomes「thinking」on the switch label, and the state is written out beside it.
• Two rows go. room.title offers Flash only (the Pro row is removed, the setting collapses to a single checked radio); the prop master's switch folds into its list.
• Usage › on a group header opens the Usage tab scrolled to that group's By-function row — a hash the usage loader already knows how to read is enough; no new endpoint.
• Docs: this page flips to BUILT in the ship commit; usage-dashboard §3 gains a one-line pointer here; the console-redesign Models section gets a「superseded by」note.

5 · What shipped beyond the mock — the owner's build rules

• fp is split in the ledger — fp.open (the opening staging, _open_stage) · fp.mid (the per-turn director, _f_engine + the two drain sites). staging aliases to fp.mid; a row booked as plain fp before the split stays fp — open or mid was never recorded, so it is never guessed. The mid-chat lead row reads Code, not Mechanical.
• V4 Pro is discontinued at 12:00 Beijing on 2026-09-14 (DeepSeek's site, 09-10): from that hour a request to the Pro id is routed to V4.1 Flash and billed at Flash's price. A dated row again (cost.PRO_ROUTED_AT): the Pro row prices on the Flash sheet from the hour, so a Pro-named seat — the floor producer's v1p arm, a fixed constant, a room saved on Pro — bills what DeepSeek bills; the 08-16 Pro row joined the superseded defaults. The greyed Pro rows in the console keep their place for V4.1 Pro.
• The Flash series is cheaper from 12:00 Beijing on 2026-09-10 (DeepSeek's notice, 09-09): $0.15 a million on a cache miss, $0.003 on a hit, $0.60 output, off-peak; peak stays 2×. A dated row (cost.FLASH_CUT / FLASH_CUT_AT), never an edit — the ledger prices a call when it is made, so the row flips by the clock; the 08-16 Flash row joined the superseded defaults so a saved console snapshot cannot shadow the cut. The specimen above shows the cut's figures.
• V4 Flash and V4 Flash Vision are gone; V4.1 Flash is the Flash tier and the eyes (owner, 2026-09-09 late: they retire 09-10, and V4.1 is the multimodal one). The four rows left MODELS; MODEL_RETIRED reads a room saved on any of them onto V4.1 Flash; every Flash-tier constant (the clerks, the gates, memory, translate, the title, the dispatch's SERP distiller, the probe's router, the photo eyes and the photo gate) says deepseek-v4.1-flash; the V4.1 rows carry vision, so a photo's pixels ride a V4.1 room's own line the way the Vision seat's did; the FP's v1f/v1fr arms are legacy aliases onto V4.1 (a saved room keeps working, the console offers v1x). Ink's seats and the Studio's steps follow. ⚠ Not baselined: the Flash-tier seats were not in the V4.1 exam (it swapped the Pro seats); the toolbox gate ran green on the swept build, the memory and Ink seats are owed their own read.
• V4.1 Flash on every DeepSeek seat, V4 Pro retired (2026-09-09, the exam): deepseek-v4.1-flash ± reasoning is a panel row, a floor-producer arm (v1x/v1xr), a prop-master and act-call choice, an option on every role radio (curator · Seen · Vibes · the Studio's steps) and on every Ink seat; the two seats that offered only Pro — the interrogation's mind (a constant until now, a choice since: probe_mind_model, thinking still the call's own) and the kit's authoring call — offer it too. V4 Pro keeps its rows greyed with a dashed「retired」chip — inert unless it is still the pick, never offered to a user, never the default — because DeepSeek routes every V4 Pro request to V4.1 Flash at Flash's price from 09-10; MODEL_RETIRED reads a room saved on Pro onto the V4.1 seat, and the rows become V4.1 Pro when that ships. The wire id rides one seam, cost.api_id: cost.V41_WIRE_DEFAULT — deepseek-flash, the GA name read off the API's model list at 10:54 CST on 09-10 (the beta id until then; not the guessed deepseek-v4.1-flash, which the API refuses) — while a one-token probe, hourly, says it answers; deepseek-v4-pro meanwhile — routed to V4.1 Flash by DeepSeek after the launch, so a wrong guess at the name never fails a turn; MAD_V41_FLASH_API is trusted without a probe (the beta id on dev before launch); settings and the ledger keep the canonical id.
• Sonnet 5 replaces Sonnet 4.6 on the panel (claude-sonnet-5; better and cheaper on the intro price). MODEL_RETIRED + model_now() read a room saved on 4.6 onto the Sonnet 5 seat at resume and at switch, so a Sonnet chat stays a Sonnet chat. ⚠ On Sonnet 5 an omitted thinking runs adaptive thinking (4.6 ran none) — the panel call now sends thinking: disabled explicitly, so the row's「thinking off」is true.
• Gemini 3.8 Flash replaces 3.7 on every Gemini seat (owner, 2026-09-09: the same price on every tier through 2026-12-31 and the same doubling after; the newer model): the Studio's second auditor and moderation fallback, Ink's rival editor, judge and copy seats, the dense-picture eyes. A saved Studio matrix naming 3.7 (or 3.5) reads onto 3.8 (builder.MODEL_RETIRED); the 3.7 price row stays so history prices. ⚠ Reviewers report 3.8 writes ~30% more output tokens on the same task — watch the Ink and Studio bills.
• Gemini 3.7 Flash replaced 3.5 in the Studio (builder.MODELS, the default audit_b, the moderation fallback); a saved step matrix naming 3.5 reads onto 3.7 (builder.MODEL_RETIRED).
• The price table is curated to what runs: Sonnet 5 · V4.1 Flash · V4 Pro (its own row in cost.PRICES now) · Gemini 3.8 Flash · Gemini Embedding. Opus and Haiku are off the table — cost.py still prices them, so the Claude web-search dispatch (Haiku) bills from the built-in sheet, uneditable here. Every figure shows as 1.0000; STT keeps six decimals because $0.000361 at four would read $0.0004.
• Inherit follows the default, live — the three Inherit rows (speech.empty · tool.act · kit.authoring) re-resolve the moment the Panel card's default radio moves, before the save round-trips.
• The dense-picture eyes follow the panel: when the default panel model is Flash Vision the row reads DeepSeek V4.1 Flash · tiled (a seeing room's own tiles read the picture); any other default reads Gemini 3.8 Flash — the routing table's rule, shown.
• room.title is Flash only — the endpoint offers one option; one checked radio.
• kit.authoring became a setting (owner, after the build): Inherit the chat's panel model (default, unchanged behaviour) or V4 Pro — authoring_model, POST /api/admin/settings/authoring-model, DeepSeek only like the act call; the call bills and tallies under the model it actually ran.
• Every fixed seat is a constant the server names — KIT_CLERK_MODEL was hoisted out of three inline literals for it, and the wake-door check now follows ACT_GATE_MODEL (env-movable) like the quick act gate it is.
• The two hops — a group header's Usage › opens #usage/<group> and flashes that By-function row (or the table, when the window holds no spend for the group); a By-function row's function cell opens #models/<group> and flashes the section.
• Sans for all on the tab, scoped to #sec-models; the usage tab keeps its mono.