The common analogy of new models being like faster but less fuel-efficient sports cars is wrong. Anthropic finds that each new model generation brings a step-function improvement in both capability and token processing efficiency, benefiting both customers and internal R&D.

It's counterintuitive, but using a more expensive, intelligent model like Opus 4.5 can be cheaper than smaller models. Because the smarter model is more efficient and requires fewer interactions to solve a problem, it ends up using fewer tokens overall, offsetting its higher per-token price.

The latest model releases from OpenAI (GPT-5.6) and Meta (MuseSpark 1.1) emphasize performance-per-dollar, not just peak performance. This marks a market maturation where labs realize enterprise adoption hinges on managing token budgets. Models are now being benchmarked on cost and latency, making efficiency a key battleground.

The cost to achieve a specific performance benchmark dropped from $60 per million tokens with GPT-3 in 2021 to just $0.06 with Llama 3.2-3b in 2024. This dramatic cost reduction makes sophisticated AI economically viable for a wider range of enterprise applications, shifting the focus to on-premise solutions.

In response to budget blowouts from agentic AI, enterprises are moving beyond simple adoption to active cost management. A new "token efficiency" stack is emerging, featuring tactics like model routing to cheaper alternatives (e.g., DeepSeek) and custom post-trained models to reduce reliance on expensive foundation models.

Large customers are aggressively optimizing AI spend by abandoning a one-size-fits-all frontier model approach. One software provider is saving nearly $700,000 annually by switching to a much cheaper OpenAI model for a high-volume task, signaling a market-wide shift towards cost-efficiency and model routing.

OpenAI's GPT-5.5 is more expensive per token, but a new evaluation framework is emerging. The key metric isn't raw cost, but the model's efficiency in solving a problem. This 'intelligence per dollar' reframes cost analysis around performance and compute, where more expensive models can be cheaper overall if they solve tasks more efficiently.

Focusing on token pricing is misleading. A more powerful model may be more expensive per token but significantly cheaper per task because its higher efficiency requires fewer prompts and iterations to achieve a final result. The correct way to measure cost-effectiveness is by the total cost to complete a job, not the price of the raw material.

An AlphaSense study revealed that models with a higher price-per-token, like GPT 5.6 Sol, can complete tasks for a lower total cost than cheaper Chinese models. This is because their superior efficiency requires fewer tokens to achieve a higher-quality result, making simple price comparisons misleading.