Open Source vs Frontier AI: Why Volume and Spend Tell Opposite Stories
Short answer: open-source models are winning token volume while frontier labs keep most of the spend, because the two serve different phases of the same lifecycle, discovery versus production.
Key takeaways
- On one major AI gateway, open-source models (led by DeepSeek) now process the most tokens, yet one frontier lab still captures more than half of total spend.
- On another marketplace, a cheap open model handled about 5.3 trillion tokens a week while the top frontier model handled just over 2 trillion, but the frontier model's average price per token was roughly 23x higher.
- Token volume and token spend are different metrics, and confusing them leads to wrong pricing decisions.
- A useful framing: frontier labs own discovery, open-source increasingly owns production.
- The practical play for builders is tiered routing, not picking one model for everything.
Why isn't open source hurting frontier labs yet?
The intuitive story is that cheap open models eat the expensive labs' lunch. The data complicates it. Decagon's CEO argued that frontier and open-source models are not really competitors, they are two phases of one lifecycle. Expensive frontier models prove out a new use case; as that use case matures and stabilizes, it gets handed down to a cheaper open model. Meanwhile new use cases keep arriving at the frontier. So volume shifts down-market while frontier spend barely moves.
What do the numbers actually show?
Two data points make the split concrete. On one gateway, DeepSeek surged to the top of token volume, processing roughly a third of all tokens, while a single frontier lab still accounted for more than half of total spend on the same platform. On a broader marketplace, a cheap open model processed about 5.3 trillion tokens a week versus just over 2 trillion for the most popular frontier model, yet the frontier model's average cost per token was around 23x higher (roughly $1.37 per million versus about 6 cents). Different leaders, depending on which axis you measure.
Why does the volume vs spend distinction matter for pricing?
If you benchmark your product against "what everyone is using," you will look at volume charts and conclude the cheap model has won. But if you are modeling your own cost of goods, spend is the number that hits your margin. A workload that is cheap per token can still dominate your bill if it runs constantly, and a premium model used sparingly for hard tasks can be a rounding error. You cannot infer your unit economics from someone else's volume leaderboard. You have to price your own mix.
How should builders act on the lifecycle split?
Treat model choice as a routing problem, not a loyalty decision. Use frontier models where quality is still being established: discovery, ambiguous tasks, anything customer-facing and high-stakes. Route mature, well-defined, high-volume tasks to cheaper open or lighter models once quality is proven. The two-tier economy is not a phase to wait out, it is the operating model. New arrivals, like Nvidia's Nemotron, only add options to the production tier.
The Calcaas lens: model both axes before you route
The mistake is optimizing for the wrong metric. Before you move a workload to a cheaper model, estimate the blended cost of a routing strategy: what share of calls stay on the frontier model at its premium rate, what share drop to the cheap model, and what that does to gross margin at each user tier. A 23x price gap sounds decisive until you remember the frontier model may only touch 10% of traffic. Put both the volume mix and the per-provider prices into one model and the right split becomes obvious. That side-by-side, provider-versus-provider simulation is exactly what Calcaas is built for.
One-line takeaway: volume leaderboards do not pay your bills, model your own spend across a frontier-plus-open routing mix.
Frequently asked questions
Is open-source AI cheaper than frontier models?
Per token, usually yes, often dramatically. In one comparison a frontier model's average price per token was about 23x higher than a popular open model. But total cost depends on how much of your traffic runs on each, not on the per-token price alone.
Why do frontier labs still earn more if open models process more tokens?
Because spend is price times volume. Frontier models command a large premium per token, so even at lower volume they can capture the majority of total spend. Open models win raw token counts while frontier labs keep the revenue.
What does "frontier owns discovery, open source owns production" mean?
It means expensive frontier models are best for new, unproven, or high-stakes tasks, while mature, well-defined, high-volume tasks can move to cheaper open models once quality is established. They serve different phases of the same lifecycle.
How do I decide which model to route a task to?
Estimate the blended cost and quality of a tiered strategy: keep hard or high-stakes calls on a frontier model and route mature, high-volume calls to a cheaper one, then check the effect on margin per tier before committing. Note: place the JSON-LD above inside a `<script type="application/ld+json">` tag in the page head.
More from the blog
LLM EconomicsProvider Substitution: The AI Cost-Cutting Lever Microsoft Just Pulled
Microsoft is now routing a share of Excel and Word prompts to its own MAI models instead of OpenAI and Anthropic, a direct move to cut AI costs and protect margins.
LLM EconomicsThe Token Apocalypse: Surviving Runaway AI Agent Token Costs
Short answer: AI agents multiply token consumption by looping, retrying, and chaining calls, so the fastest way to protect margins is to route cheaper work to the right model and model provider switches against your real usage before costs spiral.
LLM EconomicsThe Real Cost of AI: What Google and Amazon's Emissions Spike Signals for Your Token Margins
The advertised price per token is not the real cost of AI: Google's carbon emissions jumped 25% and Amazon's 16% in a year, a signal that the energy behind every inference call is getting more expensive, not less.
Pricing math, in your inbox.
One short note a week on AI pricing, token economics, and margin. No spam, unsubscribe anytime.