AI Cost Watch

Scruuge’s TLDR

Test the cheaper candidate, check your cache hits, and consider batching before changing a model that already works. Three dated calculations show the assumptions. No models were run for this edition.

Scruuge’s TLDR

This is the same AT&T team whose research underpins our cache-trap breakdown — now saying the quiet part at Fortune-50 scale: routing is the discipline, open models are where the easy majority of tokens go, and the hard minority stays on frontier models on purpose. The 90% headline hides the router and the fine-tuning it took; our Open Model Map translates which of those moves survive at small-firm volume.

Scruuge’s TLDR

Cheap didn't mean bad and pricier didn't mean safe — the cheapest model was excellent and the second-cheapest was the one that would have hurt you. Fit for the task is the axis, not price in either direction. That's why we run the real task instead of quoting a leaderboard.

Scruuge’s TLDR

The best model available and the model people actually pay for are two different things. Your peers already voted, with their credit cards, for good-enough-and-cheaper. You're not behind for staying put — you're behind if you never made the choice deliberately.

Scruuge’s TLDR

Deferring optimization used to be rational — the next release arrived cheaper and papered over your problems. That curve flattened. The work you were sensibly putting off (routing, context discipline, compiling repeats) just became the work.

Scruuge’s TLDR

The first hard measurement of what enterprises actually do with AI: usage grew 7× in nine months, and the biggest single category is documentation and technical writing. The paper never mentions cost — but that combination is the whole ballgame. Your highest-volume workload is also your most routable, and it's almost certainly running on whatever the default model is.

Scruuge’s TLDR

Every message re-sends the entire conversation — that's how LLMs fake memory — so your 10th message pays for the 9 before it, and every retry multiplies it. The fix isn't a bigger context window; it's keeping your desk clean: start fresh when the job changes, carry the artifact not the argument, and store answers you can look up.

Scruuge’s TLDR

Open-weights ≠ automatically cheaper. Judge a model by the finished bill, not the per-token rate: a verbose 'budget' model on a big footprint can beat a pricier frontier one. Pick open source for control, fine-tuning, and resilience — not a blanket cost win. Don't generalize DeepSeek's one serving win into a law.

Scruuge’s TLDR

You will not optimize your way to a flat AI bill — cheaper tokens get spent on more (Jevons). Spend isn't the enemy; unmeasured spend is. Cut the waste (floor), steer the rest toward where it compounds into revenue (ceiling), and measure per-outcome, not per-token.

Scruuge’s TLDR

The race stopped being 'who's smartest' and became intelligence per dollar, per second, per watt. A chip-starved lab just lapped the field on that scoreboard and gave away the blueprints — so the 'best model' you picked last quarter is probably not the one you should pay for now.

Scruuge’s TLDR

Cutting your bill is the floor — necessary, and soon table stakes for every competitor. The ceiling is where you point the frontier: the questions no one else thought to ask. We'll cut the floor honestly, then tell you when cost was never your real constraint.

Scruuge’s TLDR

Best model to plan, good-enough model to execute. You're not buying a smarter AI for the whole job — you're buying judgment for the 10% that needs it and paying wholesale for the 90% that doesn't. Route by task, never mid-session.

Scruuge’s TLDR

The token bill is a symptom; the disease is re-reasoning. Use AI to figure a repeating task out ONCE, compile the answer into deterministic automation, then run that — not the model — every time after.

Scruuge’s TLDR

You can't cut what you can't see — and roughly three in four teams can't. Uber isn't an outlier; it's what usage-based pricing does without a meter. Scruuge is the two-minute meter you run before finance finds out.

Scruuge’s TLDR

When the vendor quietly doubles its OWN cost estimate, that's not a rounding error — it's the trajectory. The meter only spins faster from here. Map your waste before the next 'quiet update.'

Scruuge’s TLDR

Hard proof the waste is real — and a warning we took to heart: switching models to save can backfire in long sessions. Switch your default, not your session.

Scruuge’s TLDR

Headline parameter counts are marketing; active parameters are the bill. A 'huge' MoE model can be cheap to run and a 'small' dense one expensive — always ask what actually fires per request.

Scruuge’s TLDR

Flat-rate AI is ending. Usage-based billing means your bill now scales with every token your team burns — including the 40–60% that's pure waste. The teams that mapped that waste first will feel this least.

Scruuge’s TLDR

When a chip-maker's own VP says compute costs more than the humans, the 'AI is basically free' era is over. Budget for it like the real line item it's become — then trim it.

Scruuge’s TLDR

The uncomfortable half of the story: a lot of AI spend isn't paying for itself. Scruuge's honest line — the problem is usually the WASTE, not the tool. Cut the tokenmaxxing before you conclude AI doesn't work.