Air Max: every AI model, working together in one answer

No single AI model is best at everything. One is sharpest at reasoning, another handles enormous documents without losing the thread, another is fastest for quick replies, another is strongest at fresh live-web research. For years the trade-off was forced on you: pick one model and live with its weak spots. Air Max removes that trade-off. It is not another model — it is a conductor that puts every model to work on the part of your request each one does best.

This article explains exactly how Air Max decides what to do with each message, when it uses one model versus several, and why combining models can produce a better answer than any of them alone — while still staying cost-efficient for the everyday questions that don't need the full orchestra.

One lead, four roles — chosen automatically

Air Max works in two modes and switches between them by itself, per message. The first is the Router. For the vast majority of messages — a quick question, a rewrite, an image request, an everyday explanation — Air Max picks the single best-fit engine and answers with it. That keeps things fast and cheap: one model, one answer, no waste.

The second is the Ensemble. When a message is genuinely hard, Air Max assembles a short pipeline of specialists, each with a defined role: a lead that drives the task, optional specialists that draft or reshape, a critic that verifies the work, and a synthesis step that folds the feedback into one final answer. You see a single clean response — but several engines shaped it, each doing the part it is best at.

The Air Max level ladder: L1 → L30

Every message is classified into one of thirty levels, ordered from the simplest ask (L1) to the most demanding end-to-end task (L30). The level decides how much horsepower Air Max spends: the easy levels get a single fast engine so answers stay instant and cheap, and the hard levels unlock a multi-model pipeline with a dedicated verification pass. You never pick a level yourself — Air Max reads your message and lands on the right one automatically.

The ladder is why Air Max feels both fast and deep. A greeting never drags a flagship model into the room, and a legal contract review never gets fobbed off on a lightweight one. Each rung is tuned to the real difficulty of the work.

Air Max vs a single model: how it benchmarks

The fair way to judge Air Max is not against one model on one test, but across the mix of work people actually do — quick questions, writing, coding, analysis and long research all in the same day. Because Air Max always sends each of those to the engine that is strongest at it, and then verifies the hard ones, it behaves like the best-case pick every time instead of one model's average across everything.

The scores below are directional — a way to picture the trade-off, not lab numbers. They rate the everyday experience across ten common jobs on a 0–100 scale: routing a single flagship model at everything versus letting Air Max choose and verify per task.

How Air Max verifies every answer

On every ensemble level, the answer is checked before it reaches you. Once the lead (and any specialist) has produced a draft, Air Max hands it to a dedicated critic — Claude Opus 5 — whose only job is to review, never to rewrite. Opus returns structured feedback: factual errors, missing angles, weak reasoning, edge cases and gaps. Because the critic is a different model from the one that wrote the draft, it catches the blind spots the author could not see in its own work.

The feedback then goes to a final synthesis step, where the lead model applies the critique and delivers the answer in one consistent voice. So a hard question doesn't just get written — it gets written, independently reviewed, and rewritten to fix what the review found. That draft → verify → finalize loop is where Air Max's quality edge comes from.

Powerful, but still cost-efficient

The instinct with a 'use every model' feature is to assume it must be expensive. Air Max is designed to be the opposite for normal use. Because most messages land on the low levels and route to a single, appropriately-sized engine, everyday questions cost about what they would on a mid-tier model — not a flagship. The multi-model pipeline, which is the pricier path, only fires when the level genuinely warrants it.

There is no flat Air Max fee. You are charged the real cost of exactly the stages that ran, each priced at its own engine's rate and summed — so a greeting routed to a single fast model is about 1 credit, while a full five-model mega-task tops out around 60. When several models work on one answer, yes the cost is higher than a single quick reply — but only in proportion to the extra work, and only when the task actually needed it. That is far cheaper than pricing every token at the most expensive engine in the chain.

Air Max is also the default engine everywhere — chat, agents, tools and creations all run through it unless you deliberately pick a specific model. So the cost-matching happens automatically: pick 'deep research' style once, ask a big question, then fire small follow-ups, and Air Max quietly drops those follow-ups down to a cheap fast engine instead of keeping the expensive one running. You get flagship depth when it matters and pay flagship prices only then.

Connectors and plugins: Air Max working inside your apps

Air Max is not limited to what it can say — it can also act inside the apps you already use. Connections are handled through Arcade, a hosted tool layer that manages the OAuth handshake and permission scopes for each app, so AIR never stores your passwords and every action runs under your own account with your own access.

You connect apps once from the Apps screen (previously called Connectors). From then on any Super, agent or mission can read and act through them — reading a Gmail thread, checking your Google or Outlook calendar, pulling a Drive or Sheets file, posting to Slack, filing an issue, or drafting a reply. Every tool call is written to a receipt you can inspect afterwards, and anything that changes data outside AIR passes through an approval gate first.

Powers are the per-agent view of the same system: each agent only gets the tools you granted it, so a research agent can read and search while a publishing agent is the only one allowed to post. Nothing runs silently, and nothing runs with more access than you gave it.