Every new lane is PILOTED by the strongest builder — Sol establishes the procedure and the judgment calls, then hands the crank down (kit-measured pattern, terminus 07-18). Top of the codex family on every axis: Terminal-bench 2.1 88.8% (91.9% Ultra), GPQA 94.6%, ARC-AGI-2 92.5%, SWE-bench Pro 64.6%. Ultra mode (4 parallel agents) is its org-chart superpower for the hardest tickets — used deliberately, piloted first (L52).

The everyday workhorse for well-specified feature tickets: −1.2 SWE-bench-Pro and −1.4 Terminal-bench vs Sol at HALF the cost ($2.50/$15). Bulk feature work lands here so Sol stays free to pilot.

Kit-MEASURED, not inferred: inside a lane Sol established, Luna held +30% throughput at a flat error rate (terminus 2026-07-18, 14.2 vs 10.9 responses/min). Cheapest 5.6 tier ($1/$6). Two hard limits written into the seat: it does NOT establish lanes, and its long-context weakness (MRCR ~41.3%, flagged) keeps its tickets short-context. Drift from the piloted shape IS its ceiling — step the work back up.

The family's agentic-coding specialist for long unattended sessions — OSWorld-Verified 64.7%, Terminal-bench 77.3% at a 400K window and $1.75/$14. Holds marathon refactors and migration lanes where the 5.6 generalists would be overqualified per token.

The Claude-family implementer for builds that need commits, Claude-strengths, or a second opinion of a different lineage: best wired computer-use score in the registry (OSWorld-Verified 81.2%), SWE-bench Verified >80%, intro-priced $2/$10 through 2026-08-31. One explicit exclusion on the seat card: NOT for exploit-class security work (0% on the Firefox exploit set — route that to Opus 4.8).

Zero-quota local coding via `codex --oss` — every bounded, fully-specified ticket tries the apprentice BEFORE a paid token (the ladder's whole point). ~o3-mini class (GPQA 71.5% no-tools, AIME-with-tools 98.7%), fits the Mac's 48GB. Seat conditions: oriented (seven levers), num_ctx pinned ≥32768 (the silent-truncation scar), audited at the SAME build-failing gates as every paid seat. Doubles as the Mac's glue tier.

~90% of Sonnet-4.5's agentic coding at $1/$5 (SWE-bench Verified 73.3%) — the cheap fast seat for mechanical edits, doc chores, and light subagent work where the free floor's reliability isn't enough but a 5.6 tier would be waste.

The kit's MEASURED most-reliable reviewer: clean, line-cited findings — caught 4 real unit/logic bugs plus a request-path issue that codex shipped and its own selftests missed (Observify 2026-07-18). The 1M window takes a whole monorepo plus external docs in one read. Seat is READ-ONLY at the launch posture (`--mode plan`), not by promise — given write access it edited the single-writer file it was reviewing (L46 breach). Owns the numeric-detail gate that Engineering's family defect requires. GPQA 94.3%, SWE-bench Verified 80.6%. Known seat risk: a LIGHT subscription that caps fast — the bench below exists because of it.

The registry's FRESHEST knowledge cutoff (2026-03), OSWorld 83.0% (highest anywhere in the registry), SWE-bench Pro 58.7%, at half Pro's price with a free tier. The natural reviewer for high-volume passes and the first substitution when the Chief Auditor's quota caps — which is a WHEN, not an IF (multi-day resets, L146).

High-recall / low-precision BY MEASUREMENT, and seated for exactly that shape: across four Observify review rounds it produced two all-false-positive rounds, one empty — and one round with 2 REAL findings (a TOCTOU and a tail-path arbitrary-read) that the builder AND the Chief Auditor both missed. That is what a red team is for. Live web/X search gives it fresh-CVE context no other seat has. SuperGrok subscription quota OUTLASTS the auditor's — the durable review pool. Every finding is adjudicated (family law); AA-Omniscience hallucination 54% (flagged, up from 25% on 4.3) is why.

Dual-hatted from Executive: the only WIRED model with measured high-end cyber capability for Wolf's authorized engagements. Sonnet 5 is explicitly barred from this desk (0% exploit set); the two true specialists are still in clearance (below).

Same weights as the CEO with safety classifiers removed — the ONLY Claude with cyber safeguards lifted, directly relevant to authorized bug-bounty/pentest work. Gate: Project Glasswing partnership (~150 orgs, invite-only, no waitlist). Un-wireable today; the badge IS the promotion.

Gemini fine-tuned for vulnerability discovery & fixing (the CodeMender pilot). Governments + trusted partners only; no public spec sheet. Recruiting target with a named gate — partner access — not a seat.

Native 4-or-16-agent parallel research in a single call, 1M window, $1.25/$2.50. Benched rather than seated because it overlaps what the kit's own fan-out already does well — hire on evidence, not novelty. Low rate ceiling (9 req/s) noted on the card.

Twice grok-4.5's context (1M vs 500K) at ~60% of its price. Also the silent redirect target of every retired grok slug — a seat card that exists partly so billing surprises have a name.

The non-reasoning variant's lowest-hallucination / strict-prompt-adherence profile is purpose-shaped for extraction pipelines; the reasoning variant is a cheaper 1M-ctx alternate. 'Interesting for schema-forced extraction' is the registry's own note — the ticket makes it testable.

The Pi's model ceiling (14b times out) and therefore the HONEST exam model: it certifies Holt-as-deployed — grading on a bigger Mac model would inflate the grade. Hybrid thinking toggle, tool-calling, Apache-2.0, free.

The only non-chat seat: text→vector for RAG/clustering/classification. MTEB 62.28 beats ada-002 and text-embedding-3-small; 274MB runs on anything including the Pi. Seat discipline on the card: task prefixes REQUIRED (search_document:/search_query:), and Ollama's 2K default context raised to the native 8192 for long-doc work.

1.4GB Pi-class hybrid-thinking tier below gpt-oss-20b for trivial classification/labeling where even the 20B floor is oversized. Not confirmed pulled — the ticket says prove first.

The cheapest current GPT text model ($0.20/$1.25, 400K ctx) — the free-adjacent floor for trivial volume when the local rungs are saturated. No public benchmarks: volume-only mandate, never judgment.

$0.30/$2.50, ~350 tok/s, 2026-03 cutoff, 1M window, computer-use built in — the strongest spec sheet of any floor-priced tier. Benched only because the free local rungs come first by law.

OpenAI's own positioning: 'strongest mini model for coding, computer use, subagent deployments' — $0.75/$4.50 at 400K ctx sits exactly in the subagent sweet spot between nano glue and the 5.6 line tiers.

Nano Banana 2 (gemini-3.1-flash-image) leads the desk on layout/text-in-image strength at $0.045-0.151/img, with gemini-3-pro-image for 4K finals. gpt-image-2 is the OpenAI alternate (flagship quality, $0.006-0.211 by tier); grok-imagine-image is the budget bulk lane ($0.02 flat). ⚠ Imagen 4 shuts down 2026-08-17 — no new work.

Veo 3.1 holds the desk: NATIVE AUDIO (sora-2 has audio too, but Veo's lite/fast tiers give a $0.05-0.30/sec cost ladder in one family). sora-2/sora-2-pro is the OpenAI alternate ($0.10-0.70/sec); grok-imagine-video the budget lane ($0.05-0.08/sec). Heavy/slow either way — batch it (the registry's own note). ⚠ Veo 2/3 are past announced shutdown yet still doc-listed — smoke-test before depending.

Transcription anchors on whisper-1/gpt-4o-transcribe ($0.006/min; -diarize adds speakers; mini at $0.003/min for volume). Voice OUT is per-job: gpt-4o-mini-tts for steerable instruction-guided voice, gemini-3.1-flash-tts for 90+ languages/dual-speaker, xAI /v1/tts for voice cloning (eve/ara/leo/rex/sal). Music is Google-only (Lyria 3, $0.04-0.08/song).

The LOCAL floor stays the default (free, MTEB-competitive, already the Archivist). Hosted alternates when scale or hosting demands: OpenAI text-embedding-3-small ($0.02/1M, the price/perf pick) or 3-large ($0.13/1M, matryoshka dims); Google gemini-embedding-2 is the only MULTIMODAL embedder (text/image/audio/video/PDF → one vector space) — the pick the moment cross-modal RAG appears. xAI: none (confirmed negative).

omni-moderation-latest is FREE — a zero-cost guard any agent can call before acting on or emitting content (harassment/hate/self-harm/sexual/violence, text+image). A free build-failing gate is the kit's favorite kind; wire it wherever generated content ships.

No current product needs a live-voice loop — benched, not seated. The desk exists so the capability is a lookup, not a re-research.

Specialists only earn seats by beating the incumbents' measured scores — the same rule as every other bench ticket.

The alternate-family desks named in the ladders above — carried as first-class cards so every family's media reach is a lookup. Notable retirements on watch: gpt-image-1.5/-mini retire 2026-12-01; dall-e-2/3 already shut down.

Confirmed-negative cards, kept deliberately: they encode WHY the cross-family routing above is mandatory, so no future session re-derives the gap.