The era of throwing ever-larger datasets and compute budgets at AI models may be reaching a turning point. According to a CNBC report, both OpenAI and Anthropic are recalibrating their development strategies, moving away from what the industry has informally dubbed "tokenmaxxing" -- the pursuit of raw scale through massive token consumption -- toward building models that deliver more output per dollar spent.
The shift reflects a growing tension in the AI industry between ambition and economics. Training costs remain staggering, and investors are asking harder questions about when returns will materialize. For companies like Anthropic, which recently reached a valuation approaching $1 trillion, the pressure to demonstrate sustainable unit economics is intensifying even as the race for capability continues.
What Is Tokenmaxxing and Why Is It Losing Favor?
Tokenmaxxing refers to the strategy of scaling model training by processing as many tokens as possible, on the assumption that more data exposure translates directly to better performance. For several years, this approach dominated AI lab thinking, underpinned by scaling laws that suggested predictable gains from larger models trained on more data. The problem is that those gains are becoming harder and more expensive to achieve, and the marginal improvements are shrinking relative to cost.
Key Facts
- Both OpenAI and Anthropic are reported to be prioritizing efficiency over raw scale in new model development cycles.
- Training a frontier model can cost hundreds of millions of dollars, with inference costs adding up rapidly at scale.
- Efficiency-focused approaches include better data curation, architectural improvements, and inference optimization.
- The trend mirrors a broader industry pattern seen in cloud computing, where cost-per-unit became a key competitive axis.
- Smaller, faster models are increasingly competitive with larger ones on many real-world benchmarks.
The economics are straightforward. As AI deployment moves from research demos to enterprise products, inference costs -- what it costs to actually run the model for users -- become as important as training costs. A model that is cheaper to run at scale can undercut competitors even if it scores slightly lower on benchmarks. This is already shaping product decisions across the industry, and it is one reason some AI startups have found meaningful savings by optimizing how they route requests between providers.
The question is no longer just how capable the model is. It is how capable the model is for a given cost, and that changes the entire optimization target.Industry analyst commentary via CNBC
What This Means for Anthropic's Strategy
For Anthropic specifically, the efficiency turn aligns with several observable patterns. Anthropic has been gaining ground in enterprise AI adoption, a market segment where reliability and cost predictability matter as much as raw performance. Business customers running high-volume workflows cannot absorb unpredictable inference costs, and they tend to favor providers that can offer consistent performance at manageable price points.
The company has also been expanding its policy and non-technical teams, suggesting a focus on sustainable, governed deployment rather than pure capability advancement at any cost. That broader organizational posture fits with a move toward efficiency as a core engineering value, not just a secondary concern after capability targets are met.
There is also a talent dimension worth noting. As the industry matures, the engineers who specialize in systems optimization, quantization, and efficient inference architectures are becoming more sought after. The competition for that kind of expertise is distinct from the earlier race for researchers focused purely on scaling experiments.
None of this means that capability development stops. Both companies are still investing heavily in next-generation models, and the competitive pressure between labs has not eased. But the framing has changed. Efficiency is no longer treated as a constraint to work around -- it is becoming a design goal in its own right. For an industry that has spent years chasing the next order-of-magnitude improvement in scale, that is a meaningful shift in priorities, and one that could reshape which labs are best positioned for the next phase of AI deployment.
