BRAINS
Live · ONE · IQ · Code · Velocity · IQ/$ · Peak Fable 5 · PT
A brain has reached the 10/10 ceiling - the anchor is due to be re-based upward.
| # | Brain | Status | ONE $ | IQ | Code ⚠ | Velocity | IQ/$ | Delivery 24h | 24h IQ | Public index |
|---|
Cloud rows ride the Portal rail with exact route receipts. The GPT-6 Astra row is generated on the operator’s own Codex seat (ChatGPT Pro, not a rail route) for the same slate each run and scored by the same judges; since 2026-09-10 it is ranked in ONE and priced for IQ/$ at the vendor’s own API list for the same model ($10/$50 per 1M, Codex’s harness tokens measured each run and excluded); no floor probe; an hour without a generation reads «not measured». When current ONE exists, it fixes the rank. ★ appears only when the quality lead clears the live noise band. Full methodology →
ONERank with ties (v21, the shape of Chatbot Arena, Arena-Hard and LiveBench): ONE = 100 × recent quality ÷ the leader’s, where quality = 0.8 IQ + 0.2 Code over the last 3 runs. A brain’s place counts only the brains whose quality interval lies entirely above its own (interval = 0.8 × the IQ window’s 95 % half-width); brains whose intervals overlap share a place, marked =. Among tied brains the cheaper (IQ/$) and faster one is listed first — economy and speed break ties, they never outvote quality. Nothing collapses to zero: the v11–v20 gate that printed ONE 0 for three frontier brains at IQ 6.6–7.0 is retired. The ★ MAX STAKES badge marks the separate max-stakes lane pick: highest recent dependable quality, for work where quality outranks cost and time.
The anchor10/10 = Peak Fable 5, Jun 9-12 2026 (the operator’s reference window), a declared composite ceiling. Since v19 (2026-09-04) the 10 means every communication criterion yes and a clean floor on the operator’s own shapes — the declared Peak Fable 5 experience of his communication work. When a brain sustains the ceiling, the anchor is re-based upward.
IQCommunication quality on the operator’s real shapes, measured every run (v22). Every brain works under the operator’s own house law, given to each card in the same system prompt: whole connected sentences, no em dashes, every checkable fact from the brief and nothing invented, no orders or planted negations aimed at the reader, the recipient’s specific situation named, one ask with a proposed date, every required cell present. Until v21 the board tested the bare brain from a two-sentence prompt, so the default habits that the operator’s rules normally remove (em dashes, staccato, imperatives) read as low IQ while the same brains under the same rules did his work well — the board now measures the brain as he uses it. Every run (twice a day) every brain writes the SAME four status prompts (an investor follow-up EN, a cofounder boundary message RU, a warm-circle letter after a silence RU, a companion bot’s first touch RU), on fictional composite recipients; the full run (once a day) adds two rotating breadth shapes from the other ten. First the deterministic floor (language, length, em dashes, orders at the reader, planted negations, leaks, rhythm, required cells, numbers not in the brief); then three judge seats from three model lineages, never the writer’s own, grade twenty-three anchored criteria with a verbatim quote before every verdict; splits are published. Score = 0.45 feeling + 0.25 meaning + 0.20 precision + 0.10 craft, then caps; length never rewarded. IQ = the fixed panel’s mean over the last six runs with its 95 % interval; STATUS compares each fixed prompt with the brain’s own trailing seven days on that prompt. A pairwise arena against a frozen house-brain reference (v21) remains available as an opt-in mode in the drawer methodology, not as the headline.
CodeThe operator’s code work, its own column: twelve base canaries + twelve hard functions modeled on portal-hosted semantics (Python and JavaScript, hidden tests in an isolated judge), eight per full run (once a day), latest score per task over its last three measured runs, 0.5 base + 0.5 hard. The composite 0.8 × IQ + 0.2 × Code is what ONE and the badge rank on. The v18 seeded business tier, IFEval-mini (once per release), the ten-probe floor (a health check every run) and the public index are shown, not weighted. ⚠ Saturated since v19: twenty of twenty-four canaries score ≥9.8 for every card; the column is a tie above 9.2 until a harder real-repository tier lands.
Velocity + deliverySpeed on the identical real tasks - the median wall time of one business or coding task as configured, thinking included, net of our own gateway's queue wait (10 = 4 s), not bare model inference. Delivery % counts the vendor's errors, retries and classifier reroutes over 24h; hours our own rail failed the call are shown as «our lane» and count against nobody's brain (v19.2, brain not rail).
IQ/$How much IQ a dollar buys on the identical measured workload (the full work run, once a day: two comms deliverables + twelve business tasks + eight coding tasks — the brain’s own generations, never the judge seats and, since v19.4, never the floor health probe), indexed to Fable 5.1 1M Max = 1.0×, at dated public list rates - not any provider's bill; the Astra row is priced at the vendor’s own API list for the same model with Codex’s harness tokens measured and excluded. Clean runs only, so a cheap failure can never top the column. Not a financial return.
10the anchor ceiling — both tiers clean; sustained 10 re-bases the anchor
8.5+important-work candidate; 9+ top band
7.5-8.5usable; verify before stakes rise
<7.5compare another brain; <6.5 below the work band
Embeddings
Portal's production embedding stack vs a commercial vendor on identical private fixtures - paired nDCG@10 with 95% bootstrap intervals and pre-registered verdict floors. Full methodology →
| Lane | n | Portal nDCG@10 | Voyage nDCG@10 | Δ 95% CI | Recall P/V | Verdict |
|---|
Portal-run disclosed-method benchmark. Portal AI Inc. designs and runs this comparison; Portal's own commercial route appears on the board (more than one today) and Portal has an interest in the results. Point-in-time observations - “IQ” is Portal's operational composite, not a standardized scientific intelligence test - with no warranty and no recommendation to buy or use any model. Portal is not affiliated with, sponsored by, or endorsed by any provider named here; names and marks identify the products tested (IQ/$ at list prices dated). Prompts, responses, per-item evidence, and run records are retained privately and are not published - the full methodology explains every number, formula, route, and what stays private and why. Public index = Artificial Analysis Intelligence Index v4.1.1 (artificialanalysis.ai), dated manual pins shown for context with attribution; it never enters IQ, Code, the composite or ONE. ONE Watch wires My Agent / My Telegram / My Cursor / MyX to Portal band pings - not live vendor status.
Waiting for /brains.json. Static fallback is visible.