Tokenmaxxing & the Cache Trap.

You’re paying first-class for coach seats.

When developers reach for an AI tool, they pick the premium model 85% of the time.

It feels safe. But AT&T’s analysis found that only 8–16% of tasks actually need the premium model to get a correct answer. The rest — the cheaper model handles them identically.

They have a name for the unmanaged habit of always reaching for the dear model: tokenmaxxing. It’s the AI-bill version of buying a first-class ticket for the 45-minute hop. And at scale, it’s a line item that gets noticed: on their numbers, a heavy enterprise developer wastes on the order of $1,800/month on tasks a cheaper model would have done the same.

A simple router proves the point.

AT&T trained a lightweight classifier on 3,601 real tasks to spot when premium was needed.

It predicts, per request, whether the premium model is actually required. The results (their figures):

Three times less premium usage, basically no quality loss. So far this is exactly what a cost advisor would tell you: most of your premium spend is waste.

The trap: in a long session, switching can cost more.

Here’s the part most “just use a cheaper model” advice skips.

It’s the reason Scruuge is putting it front and center. Modern LLM APIs cache your context. In a long, multi-turn session (a coding agent, a research thread), the model has already read your 200k–700k tokens of context, and re-reads them at roughly 10× cheaper than the first time.

The moment you switch models mid-session, that cache is gone. The new model has never seen your conversation — it re-reads everything at the expensive fresh rate. AT&T’s simulation of a 200-turn session is blunt about it:

Read that twice. Without cache-awareness, “intelligent” routing is the dumbeststrategy on the board — you’d be better off doing nothing. The fix is to only switch when the quality gain outweighs the cache-dump penalty.

What a small team should actually do.

This is where Scruuge’s verdict and this research agree — with one important caveat:

• Switch your default, not your session. The savings come from choosing the right cheaper model at the startof a session and staying on it — not from hopping models mid-conversation. Scruuge’s “switch from X to Y” means change your default; it does not mean switch halfway through a thread.
• One-off requests?(chatbots, classification, batch jobs) Cache barely matters — pick the cheapest model that’s good enough and move on. This is the bulk of Scruuge’s wins.
• Long multi-turn sessions? (coding agents, research) Pick one good-enough model for the whole session. The cache is worth more than the model preference on most turns.
• High volume + predictable?The cheap tier on your own hardware can approach $0 per token — AT&T’s most aggressive lever.

None of this changes the headline: most teams are tokenmaxxing. It just changes how you claw the money back without quietly setting fire to it on the way.