Frontier AI capability is converging fast, and the pricing is following. When multiple models cluster near the top of every benchmark while their token costs keep falling, "we use the best AI" stops being a competitive advantage and starts being a table stake.
What Happened
Three storylines are colliding at once. First, OpenAI has been aggressively resetting frontier pricing, signaling that the cost of intelligence itself is in freefall. Second, a mystery model recently topped AI leaderboards without anyone knowing who built it, a pointed reminder that moats around raw model quality are shrinking fast. Third, small, specialized models are increasingly beating the giants on the specific tasks that actually matter to businesses.
The platform data backs this up. Look at where top-tier competency scores now sit alongside their token costs (blended input/output rate per 1M tokens):
- Grok 4.6, competency 97/100, $3/1M blended (input $2, output $6)
- Claude Opus 5, competency 97/100, $10/1M blended (input $5, output $25)
- Veo 3, competency 98/100, $4.50/1M blended (input $4.50, output $13.50)
- Cartesia Sonic, competency 100/100, $50/1M blended (input $50, output $150)
Three of those four models score 97 or 98 out of 100 and cost between $3 and $10 per million tokens. The outlier is Cartesia Sonic: a perfect competency score at a price point 5x to 17x higher than its peers. That premium is not for general intelligence, it is for a highly specialized audio capability. That is the whole story in one data point.
Why It Matters
When the intelligence layer commoditizes, the value stack shifts. Companies that want to own their outcomes are already moving to own their data and their inference pipelines, not just rent a frontier API. The businesses that will look smart in three years are the ones building proprietary data flywheels, vertical-specific fine-tunes, and distribution advantages, not the ones who picked the "best" general model in 2024.
The Cartesia Sonic pricing is instructive: domain specialization still commands a real premium. A model that does one thing exceptionally well, in a context where general models fall short, can charge 5x to 17x more than a near-equivalent general competitor. That is where pricing power lives now.
What To Do
- Audit your AI spend by task, not by model. If a $3/1M model handles 80% of your workload, you are overpaying for the other 20%.
- Identify where you have proprietary data. That is your actual moat. A fine-tuned smaller model on your data will outperform a general frontier model on your use case, and cost less.
- Watch the specialist tier. Where AI is starting to displace roles, it is almost always a focused tool, not a general one. Specialists are where the next pricing power concentrates.
- Don't over-index on benchmark rankings. A mystery model topping a leaderboard this week means the rankings are a lagging indicator. Build for workflow fit, not benchmark position.
FAQ
Q: If frontier models are all roughly equal, why do prices still vary so much? A: Pricing reflects more than capability scores. It accounts for context window size, latency, reliability SLAs, and the provider's cost structure. Grok 4.6 at $3/1M and Claude Opus 5 at $10/1M both score 97/100 on competency, the gap is in what else you are buying with that token.
Q: Is a specialized model always worth the premium? A: Only if the task genuinely requires it. Cartesia Sonic's $50/1M blended cost is justified for production audio use cases where general models underperform. For text summarization or classification, paying that premium would be wasteful.
Q: What does "owning your intelligence" actually mean in practice? A: It means controlling the data your models train or fine-tune on, running inference in your own environment where feasible, and not being fully dependent on a single API provider's pricing decisions. It is less about the model and more about the pipeline around it.
Q: Should I wait for prices to fall further before committing to an AI stack? A: Waiting is a strategy with a cost. The companies building proprietary data assets and workflows now will have a compounding advantage. The price of tokens will keep falling; the price of being 18 months behind on data collection will not.
When every model scores 97 out of 100, the model is no longer the product — your data and your distribution are.
Hiero editorial
Bottom Line
Frontier AI intelligence is becoming a utility, competency scores are converging, prices are falling, and anonymous models can top leaderboards overnight. The durable business advantage is no longer which model you use; it is the proprietary data, domain depth, and distribution you wrap around it. Cartesia Sonic charging $50/1M for a perfect-score specialized capability while general models cluster at $3 to $10/1M is the clearest signal yet: specialization is where pricing power lives from here.