The following will be fleshed out much more comprehensively in Sapere Aude’s inaugural post, stay tuned…
As of September 2026, we are seeing two tiers of model intelligence emerging:
Tier I) Frontier Capability inelastic to cost, competing on the race to RSI-ASI (ie the creation of the “digital deity”).
Tier II) Commoditized, Task-Based Intelligence Competing on Cost Efficiency. Chinese open-source is reigning superior on this front. Enterprise token-spending trends on OpenRouter are most indicative of this (e.g. openrouter.ai/rankings).
Consider a Cost x Capability curve where at each point every model will compete on cost of intelligence per viable output at each respective tier level.
Live from Theseus Capital’s Macro Insights Lab · hover to rank what each price buys, click to pin Theseus Capital
For any given model company to create enduring value they need to win on either one of the following two dimensions:
- Frontier Capability: Astra & Fable-class models. The competition in this domain creates new scientific discoveries, novel breakthroughs, increasingly complex tasks at higher levels of autonomy. Demonstrable progress towards advanced computer use (news.ycombinator.com) and Recursive Self Improvement (arxiv.org/abs/2607.04277). A real question here is if this happens, then will everything else be rendered obsolete? This is pursuing the end game, the singularity. This is the self-professed model to end all models, triggering a fast takeoff of AI perpetually improving successive AIs. Another consideration here is if these can theoretically be distilled or replicated or even beaten by Chinese labs…
- Cost-Efficiency Per Token Relative to Capability Tier: assuming all models converge to a similar core set of capabilities, they will commoditize on price. Whoever produces the cheapest and most efficient models (per MTok) will win. It will be difficult for labs to compete with open source Chinese on this dimension. Kimi, GLM, DeepSeek, etc. For any given task that is not creation of novel knowledge (the majority of the economy), these models will be heavily commodified. They will compete on the basis of price to deliver a reliable, economically accepted outcome. Given our priors, it looks like as AI diffuses in the economy, open source will capture the bulk of this value creation.
The winner will want to dominate both tiers, but it’s unlikely.
The question is: can frontier labs win on frontier models?
Astra was impressive but I guarantee you in a few weeks the Chinese labs will produce similar results.
I think the capex of big labs is entirely predicated on frontier intelligence at this point. They are trying to create the one model to rule them all. RSI-ASI. Wall Street is underwriting under that assumption. If this doesn’t happen within a reasonable timeframe, ROI cannot justify the capex.