The race for AI supremacy isn't just about who can build the smartest model anymore; it's about who can build the smartest model for the least amount of money and compute.
The Simple Version
Imagine you are hiring a contractor to renovate your kitchen. You have two options: one contractor is very cheap but uses a massive amount of materials to get the job done, while another is more expensive but is incredibly precise and uses very little waste.
In the AI world, the "price" is the cost per token, and the "waste" is the number of tokens the model uses to solve a problem. This is what researchers call the Pareto frontier—the sweet spot where you get the highest intelligence for the lowest possible cost and effort.
As Caleb Writes Code explains, we are seeing a massive drop in the cost of intelligence, but that doesn't mean every model is getting better at using its resources efficiently.
How It Works
The Three Axes of AI Progress
To truly understand how a model like Opus 5.5 performs, you can't just look at a 2D chart of intelligence versus cost. You have to look at it in 3D, adding a third axis: token efficiency.
Opus 5.5 is designed to push intelligence further than its competitors. According to the video, Opus 5.5 can push intelligence up to an index of around 53 before it starts to plateau, whereas GPT-6 Soul tops out around 47.5 and GPT-6 Luna at approximately 37.3.
Cost vs. Token Efficiency
There is a fundamental tension between these two metrics. A model might be incredibly cheap to run (cost efficient), but if it takes 10,000 tokens to write a simple email that another model could write in 500 tokens, it is not token efficient.
The video notes that while GPT-6 models like Astra dominate in token efficiency for a time, Opus 5.5 eventually takes over because it can scale intelligence higher. Opus 5.5 scales more evenly across both cost and token axes compared to the rest of the field.
Technical Deep Dive
The technical distinction between these models lies in how they navigate the Pareto frontier. In multi-objective optimization, the Pareto frontier represents the set of optimal solutions where you cannot improve one metric (like intelligence) without degrading another (like cost).
When analyzing the trajectory of these models, we see different scaling behaviors:
- Opus 5.5: Scales close to the baseline in both token and cost axes, meaning it maintains a balanced efficiency profile even as intelligence increases.
- GPT-6 Luna/Soul: These models show high token efficiency early on, but they hit an "intelligence ceiling" much sooner than Anthropic's latest offering.
- GPT-6 Astra: This model behaves more like Anthropic's older Fable 5.1, prioritizing token efficiency over the broad cost-scaling seen in Opus 5.5.
For developers, this distinction is critical because of the API vs. Subscription split. In 2025, OpenAI's revenue was primarily driven by subscriptions, while Anthropic's was driven by API usage. This creates two different economic realities for the user.
In an API-driven model, token inefficiency is a direct cost passed to the developer. If a model is "token hungry," your monthly bill will skyrocket even if the price per million tokens is low. Conversely, in a subscription model (like ChatGPT Plus or Claude Pro), the provider often subsidizes the token usage within a rolling window, insulating the user from the model's inherent inefficiency.
Why This Matters
This competition is driving a "race to the bottom" in pricing, but it is simultaneously forcing labs to innovate on architectural efficiency. As companies like DeepSeek, Google (Gemini), and Meta continue to fight aggressively on the cost axis, the ability to maintain high intelligence without bloating token usage becomes a massive competitive advantage.
For the end user, this means the "cheapest" model isn't always the most economical. If you are building an agentic workflow—where models call other models in a chain—a slightly more expensive, token-efficient model will almost always be cheaper in the long run than a "cheap" model that wastes thousands of tokens on every step of the process.