Tokenmaxxing made perfect sense as an early-stage adoption metric for AI models. It helped organizations push engineers to experiment with AI tools, especially if teams were skeptical or slow to adopt the technology. But usage is not the same as value.

Meta recently made a public U-turn on its 'Claudeonomics' leaderboard – scrapping a tool that had 85,000 employees competing on token consumption – following backlash that it isn’t sustainable.

Once AI adoption reaches a certain threshold, maintaining a blinkered focus on token consumption only fosters warped usage incentives, unnecessary costs, and operational risk.

Engineering leaders now need to move from measuring AI inputs to measuring software delivery outcomes.

Pilot programs have an expiration date

Tokenmaxxing becomes a problem when usage stops being a proxy for adoption and becomes the goal itself.

The engineer who burns hundreds of thousands of tokens on verbose prompts, speculative exploration, or rejected code can appear more productive than the engineer who ships a small, clean, well-reviewed patch.

Measuring token consumption in and of itself rewards busywork over judgment. It sends a message to engineers, urging them to use AI even when it adds little value, simply because its use is visible and celebrated. Conversely, thoughtful prompt use or restraint is confused for poor adoption and could lead to good engineers being penalized.

Good engineers should use AI selectively, with context, oversight, and clear intent, not simply to maximise throughput into a model. So, engineering leaders should explore how to measure AI’s impact and continuously optimize their AI use to accelerate software delivery.

Exploratory AI use has a real cost

It’s obvious that token consumption has a direct cost, but many organizations are seeing firsthand that usage-based objectives can escalate quickly. Uber reportedly exhausted its entire 2026 AI budget by April. That’s $3.4 billion in R&D investment, gone in four months.

If engineering leaders are not connecting AI spend to downstream value, they are effectively paying for movement and hoping it becomes progress. While this may happen often in experimentation, it’s not sustainable at enterprise scale.

Instead, organizations need to understand where AI spending delivers measurable improvements and where it simply creates noise.

Just as we shouldn’t increase usage for its own sake, we shouldn’t reduce usage without reason. AI spend has to map to outcomes clearly.

Security and governance can’t keep up with tokenmaxxing

There are also added security risks to tokenmaxxing, which can push AI into workflows where it doesn’t belong. Under pressure to maximize usage, engineers may paste sensitive context, credentials, customer data, or proprietary code into tools without enough scrutiny.

Without the right automated quality controls in place across the software delivery lifecycle, AI-generated code can move faster than downstream review processes can keep up. Vulnerabilities, licensing issues, or architectural inconsistencies can slip through when teams equate speed with quality. The more an organization celebrates raw usage, the harder it becomes to maintain judgement about appropriate use.

For regulated or security-conscious organizations, AI adoption must be governed, not gamified.

The AI metrics that matter

The organisations that benefit most from AI will move beyond input metrics. Instead of asking how many tokens were consumed, engineering leaders should measure outcomes such as lead time, deployment frequency, review quality, change failure rate, incident impact, and developer experience.

Organizations should also reward engineers who exercise the most effective judgment when using AI to improve their workflows, not just those unthinkingly racking up the highest consumption.

Finally, engineering leaders should focus on encouraging their teams to automate repeatable workflows where AI consistently improves delivery processes, rather than rewarding broad, undirected consumption.

Leaving leaderboards behind

Tokenmaxxing helped to usher in the broad adoption of AI we see today. But its time has passed as a primary metric for identifying and recognizing AI power users.

Now, the priority should be identifying where AI creates engineering value. That means putting the leaderboards to one side, and taking a more outcome- and value-driven approach to measurement.

Has your organization ditched its token leaderboard? Did you never have one in the first place? Let us know how you benchmark AI adoption success across your teams.