Find your question
Seven pages, each answering a different kind of question. If you don't know where to start, this table probably has your answer in it.
| You want to know | Go to |
|---|---|
| What's the market doing today, at a glance | Overview — the price map, the week's biggest movers, data status |
| What just launched, or what actually changed in price recently | Discovery — new releases since day one, and every published price move |
| What does a specific inference price index say right now | Indices — pick a family, a measure and a segment; the number and its chart follow |
| Which number is meant to be referenced by a hedge, not just watched | Indices → Settlement grade — see below |
| What would my own input/output ratio cost, not one of the five presets | Indices → the Custom blend card, below the switcher — see below |
| Why is the market cut this way — does Frontier just mean Proprietary, is the equal-model ranking actually different, where does the bill's risk concentrate | Analysis — read-once material, not a daily number |
| What does one specific model release cost, and who serves it | Models — search or scroll the published table, click a row |
| How has one release's price moved over time | Click a row on Models, then "Price history" for daily or per-run detail up to a year back |
| How a number is actually computed, or what a status/flag means | Methodology |
| The raw numbers, for your own use | API reference — every page here reads from it |
Overview
The front page, and the only one built for a five-second read rather than a specific question. The price map plots every published release by capability against price — a wide, cheap cluster low and left is the market's center of gravity; the isolated dots up and right are the expensive frontier. Below it: which releases moved the most this week, and whether today's move in the market unit value is a real price change or just a shift in which models people are using (the price-change-or-mix-change card).
Indices
The switcher at the top picks three things, and stays pinned as you scroll:
- Family — Core indices (the primary set, capability tiers ranked by tokens actually used), Settlement grade (methodology v0.4, built to be referenced by a derivative rather than just observed — see below), Equal-model tiers (the same construction as core, ranked by model count instead — a sensitivity check, not a second index to trade), or Unit value (not an index at all; see its own caption).
- Measure — Input, Output, or a blend (80/20, 50/50, 25/75 input share). These are standard scenarios everyone can cite, not a claim about your workload.
- Segment — the chips below the switcher: Broad market, a tier (capability for core/equal-model, price for settlement grade), Open weights, Proprietary.
Whatever you pick, the card below shows the number, its chart (toggle between index points, basket cost in USD per 1M tokens, and day-to-day change), and — for a core, settlement-grade or equal-model series — the basket's actual members and weights.
Core's monthly reconstitution and quarterly tier review aren't a slower version of settlement grade — they answer a different question on purpose. Core is meant to be read as a single trend: a slow cadence means a month-over-month move is (almost) never partly "the basket changed," so the number stays comparable across readings. Settlement grade trades that away deliberately, because a basket meant to back a hedge that's a month stale is a bad hedge. Read core for "what does capability cost, broadly, over time"; read settlement grade for "what would I actually reference in a contract."
Settlement grade
A second, parallel index family (methodology v0.4) built to be referenced by a derivative, not just observed: weekly reconstitution instead of monthly, an incumbency buffer so that faster cadence doesn't flap. Its tiers — Premium, Standard, Economy — rank by price rather than capability, since a hedge bucket needs its members to be price substitutes, which a capability tier isn't especially close to being (see Methodology for the measured numbers). Its own board sits above "Core, equal-model and unit value" on the Indices page, and its series carry the AIX.SG. ticker prefix. The one mechanism not exclusive to it: a member with no fresh quote from any provider for more than two days is excluded from that day's weights and disclosed rather than carried indefinitely — core does this too now, since a member gone dark is a data-integrity fact rather than a tradability one.
Custom blend
Right below the switcher, before the selected series. Nothing you pick here is stored anywhere or becomes a new index — it's a calculator, computed on the spot from the same two numbers behind every blend on the page: the segment's input cost and output cost, both already published.
- Pick a segment above as usual (Broad market, a capability tier, Open weights, Proprietary — whichever the calculator should price). The card updates to match automatically.
- Drag the Input share slider to your own ratio — say 85% input, 15% output for a context-heavy agent workload, or 30/70 for something output-heavy.
- Read the price: USD per 1M tokens, today — not index points. A blend of a dollar cost needs no reconstitution math, which is also why this works for any ratio, not just the five presets: it is exactly
alpha × input cost + (1 − alpha) × output cost, alpha being your slider position. - The second figure is how far that is from the standard 50/50 blend, so you can see whether your workload is cheaper or pricier than the "headline" number implies.
- The chart below is that same calculation run across every fixing day so far — your ratio's own curve, not an index level, so you can see it move without translating from points.
What it's for: a monthly bill estimate at your real ratio, a fast comparison of what a given capability tier costs at that ratio instead of at 50/50, or a specific, defensible number to put in a spec or a message instead of eyeballing between two presets. It answers "what would this cost me", not "is this a good index" — for that, the five standard measures are the ones worth citing, because they're the same for everyone.
Analysis
Read-once, not a daily check-in. Every card states the question it answers before it answers it. Currently: how the market's different cuts (capability by tokens, capability by model count, open vs. proprietary) compare against each other; whether the Frontier tier is secretly just "proprietary" by another name; where a bill's dollar exposure actually sits relative to where its tokens go; and the spread between segments in one number each, updated daily.
Models
The full published table — every eligible release's weighted median price, ranked by token volume. Click a row to open its detail without leaving the page: providers behind the price, its index tier memberships, a small price-history chart, and a "Price history" button through to the full record — one release's price over time, at daily or per-run resolution, up to a year back. That deeper page has no nav entry of its own; it's always reached from a row here, or directly with a release in the URL.
Discovery
The browsing page: what's new, and what actually moved, without a specific release in mind already. "What's new" lists releases first seen in the last 45 days with their price path since launch — open-weights prices typically fall as more hosts start serving them. "Recent price changes" is close to the whole published record for proprietary models, which almost never move. The Chinese labs lens filters both to a curated set: labs confirmed against a source get a plain badge, and labs we haven't verified — the company's nationality, or this specific creator tag's mapping to a real company — show "on radar" with the specific reason, never silently merged in with the confirmed set.