Vivek Gopalan: Companies with the Worst Final Margins and Paying Cutting-Edge Model Prices Often Spend the Least Time on Evaluations
Venture capitalist Vivek Gopalan stated that companies with the worst final margins and those paying cutting-edge model prices often spend the least time on evaluations.
He described the internal practices of Anthropic and OpenAI: automating the strongest internal models at the time; once the process is established, using Opus or Fable to generate evaluations based on their own trajectories, then migrating down the intelligence hierarchy until they can no longer pass. Most large companies spend innovation budgets on pilot projects without monitoring unit costs, making the cutting-edge premium invisible, but it will soon be apparent to most.
He anticipates that evaluation coverage—the proportion of token spending on real evaluations, analogous to test coverage—will appear in due diligence, closely following gross margins, especially for AI-native companies where inference constitutes a significant portion of sales costs. Two companies with the same revenue and using the same batch of models can have vastly different gross margins, with the gap mostly stemming from evaluation coverage and the synthetic environments generated by evaluations.
The feedback loop will also move upwards: each generation of new cutting-edge models first rewrites the standard answers at the top of the hierarchy, as the evaluation limits generated by the previous generation's trajectories represent the quality of that generation; then the entire line migrates downwards, with the thresholds typically higher than last time, as cheaper models have also improved. Third-party services like ValsAI have ranked models based on task costs and accuracy, placing capability, latency, and unit price on the same table.
If internal automation remains stuck on the most expensive models, unit economics will be obscured by innovation categories. Coverage turns "can we use cheaper models to accomplish the same work" into an auditable number.
In market mechanisms, this is a shift in cost accounting rules, not model releases. Buyers need to see the investment and financial officers who understand inference costs; sellers are tool companies selling evaluation platforms and synthetic environments. Funding shifts from indiscriminate cutting-edge calls to layered routing. Beneficiaries are operational teams that can use evaluations to push traffic to lower-priced models, while pressured are innovation departments that run all pilots on the most expensive interfaces. A single post does not change the price list but will change what is asked in the next round of financing.
Source: Public Information
ABAB AI Insight
Software companies use test coverage to determine whether to release versions, while model companies are still stuck on "using the best first." Gopalan applies the same discipline to token billing: first using the strongest model to get the process right, then generating exams from trajectories and handing over the passable parts to cheaper models. Opus and Fable are written as exam creators, not permanent executors. Once the innovation budget disappears, the cutting-edge premium will directly impact gross margins.
The capital path is an upgrade of the due diligence checklist. The proportion of inference in the sales costs of AI-native companies is rising, and investors need a single metric similar to test coverage. High evaluation coverage means the same revenue can be delivered using lower-tier models; low coverage equates to handing gross margins over to the price difference of model manufacturers. Synthetic environments allow evaluations to be regenerated in bulk, rather than relying on manual labeling that becomes outdated.
This is analogous to cloud computing transitioning from using only the highest configuration instances to automatically scaling down based on load; it is also like compiler optimization keeping hot paths on fast paths. The industry phase is shifting from who can tune models to who can layer tasks. Internal labs have already been doing this, while corporate pilots still default to the most expensive models.
Structural judgments belong to the transfer of pricing power. Model manufacturers price according to capability tiers; if clients lack evaluations, they can only buy the most expensive tier. The mechanism is: exams are generated by stronger models based on real trajectories, which allows for understanding the boundaries of weaker models; after new models are released, the answer keys are rewritten, and layering can then shift down along with cheaper tiers. The gross margin difference comes from routing, not from whether the same model can be purchased.
ABAB News · Cognitive Laws
Companies that always use the strongest models are not buying capability, but invisible premiums.
Evaluation coverage will enter the due diligence column next to gross margins, just like test coverage.
If the exam does not upgrade, cheaper models will not pass the previous generation's questions, no matter how much they improve.