CEOs like Mark Zuckerberg have been mocked for token-maxxing leaderboards and contests rewarding engineers for the most token usage across coding agents. One internal Meta leaderboard, "Claudeonomics," even handed out titles like "Token Legend" before being shut down. The numbers behind it are absurd: Meta employees burned through 60 trillion tokens in 30 days, a sum that would cost roughly $900M per month at Anthropic's public API rates.
Admittedly, I asked our CTO if implementing a similar reward for our engineering team was a good idea a few months ago, before token leaderboards became popular. While I recognized that token output was a suboptimal metric and could easily be gamed, it was a blunt-force instrument: objectively measurable and easy to reward. My reasoning, like that of many CEOs, was twofold:
I believe very strongly in Charlie Munger's line: "Show me the incentive and I'll show you the outcome." Sometimes, to usher in dramatic change at a company of any size, you need a short-term incentive, however flawed, to transform behaviours. I, like Zuck and others, wanted our engineering team operating at the leading edge of coding agents, and a token contest seemed like the right incentive to push them past their comfort zone.
I wanted our engineers to feel they could use these tools unconstrained by financial costs in pursuit of maximum productivity. With the CEO's explicit permission to dramatically increase token usage, an engineer would feel not just comfortable but financially justified in spending exponentially more on tooling to maximize productivity.
On our CTO's advice, I rightly decided not to pursue the token contest. The second-order effects could create too much of a distraction. But I've been monitoring the discourse on token leaderboards and felt it was worth offering rationale rather than just criticism.
Token-maxxing is an extreme example of AI transformation. Coding agents have become by far the runaway success of vertical AI agents. They don't just transform the profession of engineering, they change the way the world builds and interacts with software. When used well, they can increase the output of a top engineer by 100x. AI transformations at most companies won't include token leaderboards, but the principles behind them, and behind my ask to our CTO, apply broadly to every company's AI transformation.
Most importantly, every AI transformation needs more than exec buy-in. It needs a Tobi Lütke-style directive: "reflexive AI usage is now a baseline expectation."
Two examples from our customer base.
Exec directive is required to change behaviour. Inertia is a powerful force. People are busy and set in their ways. The status quo is easy. Across our customer base, we've seen Gigi deployed successfully when the agency CEO states: "we are changing the way we work to be AI-native." Conversely, without a top-down directive, front-line users of Gigi, who are tasked with learning a new way of working, at times opt out, become less engaged, and when asked "should we continue leaning into Gigi?" respond ambivalently. In the same way some engineers prefer to stay in control rather than lean into coding agents, functional leaders and media managers may say their current way of working is good enough. Part of this is on us. We need to show the magic of working with Gigi to every stakeholder. But this decision can't be left to front-line employees, because only a small few will be early adopters and only a small few want to put in the "extra work" to become AI-native with urgency. It's too easy for functional leaders and front-line employees to maintain the status quo. It takes an executive's directive to say: "we are doing this. We are doing this because it will lead to better outcomes for our clients, which Gigi can demonstrably show. And we are doing this because if we don't do this now, we may not have the luxury of even having this decision in the future."
Exec directive is required to absorb short-term cost increases for long-term operating efficiencies. Gigi, like many vertical AI agents, isn't unseating an incumbent. Our costs have to be rationalized against long-term workforce efficiencies, and those efficiencies will likely be borne out over several years. In no world is someone hiring Gigi, or any vertical AI agent, and immediately reducing their teams at scale. So every AI agent leads to incremental costs in the near term, and traditional budgeting processes don't accommodate that. I met with the exec and functional leaders of a large agency at POSSIBLE last week. Once the exec got more familiar with our business, he asked the functional leaders: "why aren't we adopting this more broadly across our client base?" The functional leaders cited near-term cost concerns. The exec replied: "I'll underwrite the immediate costs because I know the operational efficiencies this will bring our business over time." Traditional budgeting processes are rigid and myopic. It takes forward-thinking execs to rationalize those costs and empower their teams to incur them now, because the alternative is being forced to do so later when it's existential.
So yes, token leaderboards and token-maxxing are flawed and worthy of some criticism. But every exec of every company needs their teams thinking like token-maxxers if they're going to thrive in this new AI-first way of working.
Cherry Picked is a monthly newsletter from Adam Epstein, co-founder and CEO at Gigi, covering the AI and commerce media insights you just gotta know.
