# Overpaying for incremental intelligence

> The best AI model is often a few percent smarter than a rival and several times more expensive. That premium is not just wasted spend. It is the ceiling on how much intelligence you can afford to run.

Published: 2026-07-20. Last updated: 2026-07-20.

Anthropic's Claude Fable 5 is 11% more intelligent than xAI's Grok 4.5 on the Artificial Analysis Index. It is also 567% more expensive. The chart above puts every major lab's flagship on those same two axes, and it tells one story: intelligence has bunched up, and price has not.

## The releases stopped being about the frontier

The last two weeks made that hard to ignore. Meta shipped Muse Spark, xAI shipped Grok 4.5, and Moonshot shipped the open-weight Kimi K3, which landed at 57 on the index, effectively tied with the best models OpenAI and Anthropic sell ([Artificial Analysis](https://artificialanalysis.ai/articles/kimi-k3-achieves-3-in-the-artificial-analysis-intelligence-index-comparable-to-opus-4-8-and-gpt-5-5)). The six leading flagships now sit inside a ten-point band, 50 to 60. Ten points separate the best model in the world from the sixth-best, and on most real work you would not notice the difference.

## The premium buys almost nothing

Now look at what the premium actually buys. Blended at a standard 3:1 input-to-output mix on OpenRouter list rates, prices run from about $2 per million tokens for Muse Spark to $20 for Fable 5. Moving from Kimi K3 to Fable 5 buys you three index points and roughly triples the bill. Grok to Fable 5 buys six points at 6.7x the price. This is a diminishing-returns curve, and the habit of defaulting to the best model puts you at the steep, expensive end of it, paying a luxury markup for intelligence most tasks never call on.

## The real cost isn't the invoice, it's the throttle

Here is the part that never shows up on a per-token price sheet. Your intelligence budget is finite. A model that costs 6.7x as much doesn't just cost more per call, it means 6.7x fewer calls for the same money. The premium quietly sets a ceiling on how much intelligence you can run at all.

That ceiling is where the real money leaks. Picture an agent that watches customer activity every hour, flags the accounts showing buying intent, and drafts outreach the moment a signal appears. On the cheap tier you run it continuously. On the frontier tier, at 6.7x the cost, the same budget only buys a bulk pass once a week. By then the prospect has already bought from someone else. You didn't pay for a smarter answer. You paid for a slower one.

As companies push programmatic intelligence into every workflow, this inverts the question. The constraint stops being "which model is smartest" and becomes "how much intelligence can I afford to run around the clock." Price per token is the new bottleneck.

## The ROI question the market is now asking

This lands at an awkward moment for the premium labs. OpenAI and Anthropic both filed IPO prospectuses in June, targeting roughly $1 trillion and $965 billion respectively ([Barchart](https://www.barchart.com/story/news/3116486/openai-anthropic-ipo-what-s-confirmed-what-s-still-speculation)). Those numbers assume the frontier stays a moat you can charge a steep premium for. But the underlying economics are strained: one analysis pegs the three headline AI IPOs as needing on the order of $150 billion in combined annual earnings to justify their valuations, against roughly $20 billion in combined losses today ([Intellectia](https://intellectia.ai/blog/ai-mega-ipo-wave-2026-spacex-openai-anthropic)). If a $6 open-weight model is 95% as capable as a $20 one, the premium tier's pricing power is precisely the assumption under the most pressure.

## Make intelligence a dial

The reason teams overpay is rarely ignorance. It's friction. In most stacks the model is welded in: prompts tuned to one provider, parsing built around one format, the whole thing tested against a single API. Switching is a migration nobody wants, so the premium compounds one request at a time.

General Input removes the friction. The model is a setting, chosen per workflow. Run a customer-facing answer on the frontier and a bulk classification job on a mid-tier model. Or hand the decision to a router that sends each task to the cheapest model clearing your quality bar, and re-checks as cheaper models keep improving. Because the model never sees your credentials, which are encrypted at rest, injected at runtime, and stripped before anything reaches it, with every external call logged, swapping models is a config change, not a security review.

The winners in the next phase of this won't be whoever owns the single smartest model. They'll be whoever can deploy the most intelligence per dollar. You're paying frontier prices for a rounding error of extra intelligence. Price the increment before you buy it.

## FAQ

### Is the most intelligent AI model worth the price?

Usually not, for most tasks. The six leading flagships now sit within ten points of each other on the Artificial Analysis Intelligence Index (50 to 60), while their blended prices span 10x, from about $2 to $20 per million tokens. Moving from a near-frontier model to the top one often triples or 10x-es the bill for a few index points that most workloads never use. The frontier model is worth it for the specific tasks that genuinely need it, not as a default for every call.

### How much cheaper are non-frontier models?

A lot. Blended at a standard 3:1 input-to-output mix on OpenRouter list rates, Meta Muse Spark runs about $2 per million tokens, Grok 4.5 about $3, Kimi K3 about $6, and GPT-5.6 Sol about $11, against $20 for Claude Fable 5. Kimi K3, an open-weight model, scores 57 on the intelligence index versus Fable 5 at 60, so you keep roughly 95% of the capability at under a third of the price.

### What is the efficient frontier for AI models?

It is the set of models where you cannot get smarter without paying more, or pay less without getting less capable. A model that another model beats on both price and intelligence is dominated and sits off the frontier. Choosing well means picking the cheapest model on the frontier that still clears the quality bar for your specific task, rather than defaulting to the most capable model for everything.

### How does General Input reduce model costs?

It makes the model a per-workflow setting instead of a hardcoded foundation. You can route a customer-facing answer through a frontier model and a bulk job through a cheaper one, or hand the choice to a router that sends each task to the cheapest model clearing your quality bar and re-checks as cheaper models improve. Because the model never sees your credentials and every external call is logged, swapping models is a config change, not a security review.