Tokenmaxxing is dead, long live valuemaxxing

Over the past year, developers and organizations saw firsthand how agentic AI could accelerate workflows, and some sought to formalize increased AI usage across their engineering teams. Some leaders even encouraged developers to “use as much AI as possible” without clearly defining success metrics, guardrails or cost expectations. The term coined was tokenmaxxing: the practice of maximizing AI usage across development workflows with the conviction that heavy AI consumption would eventually lead to better outcomes.

That approach made sense during the experimentation phase. “In the absence of true metrics, organizations created usage leaderboards, which people quickly learned to game,” Neil Dhar, SVP for IBM Consulting, wrote on LinkedIn. “[U]sage soon became a proxy for value.”

But maximal usage has faced a new challenge: the rising costs of AI-assisted development without guardrails. Agentic development workflows do more than generate code. They plan, analyze repositories, call tools, test solutions and coordinate multiple steps across complex systems. As usage scales, token consumption scales with it, and some organizations are starting to ask a difficult question: what happens when we have to pay for all of this, and what did we get out of it?

As AI costs increased, many organizations responded by focusing on the most visible metric available: token consumption. New policies emerged around limiting usage, restricting newer models, shrinking context windows and reducing prompts.

The problem is that token minimization can fall into the same trap as tokenmaxxing. Both assume token consumption is the primary metric that matters. One seeks to maximize it. The other seeks to minimize it. Neither measures business outcomes.

Organizations often mistake reducing visible token consumption for reducing actual costs. Once obvious inefficiencies such as oversized tool catalogs, unnecessary payloads or stale context are removed, further reductions frequently target the information that helps AI systems succeed: task descriptions, business constraints, architectural context and other sources of meaning.

At that point, costs do not disappear; they move. Ambiguous instructions create additional reasoning, retries, tool calls, validation cycles and human rework. The organization may celebrate lower input-token counts while paying for the same complexity elsewhere in the workflow.

The goal should not be token maximization or token minimization. The goal should be valuemaxxing, a term coined recently by Marc Boroditsky, the Chief Revenue Officer (CRO) of Nebius.

Valuemaxxing moves the conversation toward direct outcomes. Instead of asking, “How many tokens did we use?” organizations should ask:

• How many tasks were completed?
• How much developer time was saved?
• How much modernization work was accelerated?
• How many vulnerabilities were resolved?
• How much rework was avoided?

These are some of the metrics that determine whether AI is creating business value and whether incremental token usage is justified. Token consumption is a cost signal, not a value metric.

This distinction becomes important as models become infrastructure or system dependencies. Access to performant models is no longer the primary differentiator. Most leading models can successfully perform complex development tasks, and high-quality open-source alternatives continue to improve. Competitive advantage is shifting away from the model itself and toward the systems built around it: context management, workflow orchestration, memory, governance, evaluation and optimization. If outcomes matter more than tokens, then systems (outputs) matter more than models (inputs).

Model orchestration is also becoming more important than model selection. IDC predicts that by 2028, 70% of leading AI-driven enterprises will dynamically manage routing across multiple models rather than relying on a single-model strategy.

For developers, AI efficiency is becoming a new engineering skill. Developers should think about AI usage the same way they think about cloud resources, databases or application performance. The objective is not to use less AI but to use AI more effectively. Good context hygiene, planning before execution and understanding tradeoffs between cost and outcome all contribute to higher-quality results.

For leaders, the responsibility is different. Leaders must create visibility into both costs and outcomes, and measure value instead of volume. Efficiency should be rewarded, and teams should have the tools to understand their impact to track outcomes.

Building a culture of AI efficiency requires sharing metrics, accountability and visibility across the organization.

Developers and leaders cannot shoulder the responsibility for AI efficiency alone. The platforms they use should also be accountable for helping them control costs and connect AI consumption to outcomes.

This philosophy is one of the reasons IBM Bob, IBM’s agentic development platform, was built the way it was. Rather than optimizing for a single model, Bob was designed to remain resilient as models, costs and enterprise AI continue to evolve. Intelligent model routing balances frontier and open-source models, so teams use the right model for the right task, rather than defaulting to the most capable (or most expensive) option.

Bob also provides the transparency organizations need to practice valuemaxxing. Administrative controls deliver budgets, governance and usage visibility, while analytics help leaders connect AI consumption to business outcomes instead of token counts alone. Workflow capabilities, skills and tool integrations reduce unnecessary work while preserving the context and signal needed to deliver high-quality results.

More than likely, yes. Over the next few years, model efficiency may improve dramatically. The cost of compute may decline. Open-source models may continue closing the gap with frontier models. New hardware, architectures and deployment approaches could fundamentally change the economics of AI.

In fact, many of these trends are already underway. The rise of open-source models and the growing importance of model routing are shifting attention from simply accessing the most powerful model to orchestrating the right model for the right task. This is one reason platforms like IBM Bob have invested in model routing from the beginning: helping organizations balance performance, cost, governance and flexibility as the model landscape continues to evolve.

But regardless of where costs go, organizations should be able to answer a basic question: did our AI investment create meaningful business value?