Welcome to Product Cocktail, where the takes are as polarizing as a shot of Fernet—but the insights come together like a perfectly crafted daiquiri.
The Shake
Eight weeks ago, I called token leaderboards AI adoption cosplay. This week, I take my victory lap.
Around that time, a Claude Code skill called "Caveman" was going viral on GitHub. Its tagline: "why use many token when few token do trick." Caveman strips Claude Code replies to the studs, using as few words (and output tokens) to convey the message. Grammar and verbose AI speak be damned.
ME FIND BUG. BUG BAD. ME FIX NOW.
In the last two months, Caveman climbed past 85K GitHub stars and transitioned from esoteric nerd joke passed around Hacker News to "company policy."
Caveman's GitHub stars, April-July 2026. The near-vertical ramp happened before I ever hit send on Issue 4 — the revolt was already underway. (Source: github.com/juliusbrussee/caveman)
According to Caveman creator, Julius Brussee, "people using caveman include developers at OpenAI, Nvidia, and GitHub." Infrastructure giant, Legrand recently told staff in an internal memo to "use 'caveman skill'" as a sanctioned way to stay under new AI budget quotas.
Caveman as company policy is the desperate reach for the tepid bottle of Pedialyte on your nightstand as the entirely predictable hangover headache of token overspending kicks in.
The Whiplash
In Q1 2026, "Token Legend" was a status symbol. Meta employees torched 73.7 trillion tokens in a single month — north of $200M at list price — with one "Cache Wizard" alone burning 281B (a cool $1.4M). Shopify tied token use to performance reviews. Amazon built an official leaderboard (because of course they did). The thesis was more tokens = more productivity. This was the de facto, half-assed vanity metric that companies chose to co-opt to measure adoption.
In Q2, the bill came due. Companies started backing off the bacchanal of big data.
Two weeks after my article, Amazon shut down their "KiroRank" leaderboard, with SVP Dave Treadwell telling staff "please don't use AI just for the sake of using AI." (LOL)
Staring down the barrel of a $300M Anthropic bill, Marc Benioff (Salesforce) mused on the All-In podcast that he wished there was an "maestro that can route [a query] to the most affordable [model] for the job."
It seems that the cache wizards have given way to token minimalists.
On Tuesday, Morning Brew ran a parody on "token mini-maxing." Token temperance has officially reached the mainstream.
What this actually means
Token leaderboards signaled an immaturity around measuring AI impact giving way to measuring activity instead. Caveman, and the litany of Finance department co-sponsored "oopsie" missives from Big Tech, signal a collective admission that we burned too much on activity, so we're moving to rationing.
Neither is a strategy. One is a vanity metric, and the other is vibe-based austerity.
The same Legrand memo mentioned above ranked "use caveman skill" last on their token thrift guidelines. The three practices above it were: 1) don't default to the most powerful model, 2) don't default to high reasoning, 3) route tasks to the right model.
Even the company that adopted a meme-as-policy ranked it dead last. These principles are all just saying "match the work to the right-sized model instead of reflexively grabbing the biggest one" — the same thing that's on Benioff's wish list. That's the discipline I said was table stakes back in May.
Issue 4 said stop counting tokens, measure outcomes. There's already a shift to outcome-based pricing in applied AI tooling (Zendesk, Intercom), with Pegasystems as the latest domino to drop, complete with product marketing squarely aimed at token-spend-triggered CFOs. In the same breath as they shuttered KiroRank, Amazon moved to track "normalised deployments" (whatever that means) to measure engineers' creation of "useful code."
Translation: "CFOs, we saw your token bill too." The vindication isn't subtle when it shows up in a vendor's landing-page copy. (Source: Pegasystems)
From the same piece on Pegasystems, Liz Miller from Constellation Research states "tokenmaxxing was a mentality of no-holds-barred experimentation driven by a fear that AI wouldn’t be adopted quickly enough."
The ones who are getting this right never needed the leaderboard or the caveman. They aren't living in a GEICO commercial — they set the right conditions and let the outcomes do the talking.
The Recipe
A token budget is not a strategy, but it’s a start.
Too much: Caveman-mode as the fix. Token austerity for its own sake — rationing employees, revoking licenses, cutting off your non-technical users to spite the invoice. Swinging the pendulum from "burn everything" to "burn nothing" is the same vibes-based management with a minus sign.
Not enough: Ownership of token cost, instead of treating it as someone else's problem. AI PMs must consider questions like model routing, context management, and output scoping.
The fix: The middle way of intentional AI deployment with defined outcomes was missed when the "adopt or die" train left the station in January 2026. The tools and the playbook are becoming clearer, the missing ingredient is organizational willpower to treat AI as a product decision, not a performance metric.
The Garnish
“It’s not actually our engineers that are driving the token consumption. It’s a lot of the non-engineers…”
-Justice Kwak, Accenture’s agentic AI strategy lead, in a leaked internal meeting
You thought that all of this AI spending was at least engineers pushing code? Think again. Accenture’s token tab is being driven by its staff using AI to convert PDFs to slides.
After pushing its clients to rapidly adopt AI, it’s finding a new way to turn the knife: advising them on “token economics” to control AI spending.
If there’s one thing you can always count on with consultants, it’s that they’ll always invent a new, creative way to get you to spend money with them.
Back in my day, we used to bill a client $250/hr for an analyst to convert PDFs to presentations.
Source: 404 Media Co
Product Cocktail
Tip Your Bartender
Send me questions, feedback, and cocktail recipes:
[email protected]
Icons made by Icongeek26 from www.flaticon.com.