All Inferred editions

Inferred 04

Model portfolios

This week - a deep dive into AI-native startups and large organizations on OpenRouter. How many models do they use, how much of the work goes to their main model, how often do they swap that main model out, and which tasks hold on to a model the longest.

Highlights

  • Portfolios grew by 40% in a year: The average AI-native startup on OpenRouter used 12 models in September, up from 9 last October. The top 10% used 28 last month.
  • One model still does two-thirds of the work: The top model handles 66% of the average large org’s requests. That’s only 2 points lower than last October, even though the average portfolio grew by about 40%.
  • Switching the main model is routine: Every month, roughly half of these orgs have a different top model than the month before. In September it was 57% of AI natives and 54% of large orgs, and about three-quarters of those switches went to a model from a different lab.
  • Specific workloads switch at different rates: For general coding, 53% of accounts changed their top model from one month to the next. For translation it was 24%, for summarization 27%, and for customer support 30%.

Portfolios keep getting bigger

Models used per month, by percentile, AI-native startups and large orgs.
Models used per month, by percentile, AI-native startups and large orgs.

Over the past year, the average AI native startup went from 9 models a month to 12, and so did the average large org, though the median stayed put around 6.

However just like startup fundraising, there is a power law in model users. The 90th percentile for AI native startups went from 18 to 28 models. The 75th percentile went from 11 to 14.

A typical org runs a handful of models and adding a new one to the portfolio is a big decision. A large minority runs dozens, sending different workloads to different models and trying most new launches as they arrive.

One model still does most of the work

Average models used per month against the request share of the top model.
Average models used per month against the request share of the top model.

More models doesn’t guarantee traffic gets spread more evenly. In September, the average AI native startup sent 68% of its requests to its top model, and the average large org sent 66%.

So while the average portfolio grew by about 40%, the main model lost only 2 points of share. That suggests a hub-and-spoke setup. One model acts as the “daily driver”, carrying two-thirds of the request load. The other ten or so models split the remaining third: evaluations, specialized tasks, fallbacks, and new launches being tested. New models get added at the edges much faster than they replace the hub.

Switching the main model is routine

Share of orgs switching their top model, and switching to one from a different lab.
Share of orgs switching their top model, and switching to one from a different lab.

In every month of the past year, between 48% and 57% of these orgs had a different top model than the month before. So for roughly half of them, the main model changes in any given month.

The increasing pace of model releases serves as the background context to this seemingly major choice. Waiting 2 or 3 months to give a new model a try as your daily driver may mean you are more than one generation behind the frontier.

The bigger shift is where orgs switch to. Last October, about two-thirds of switches went to a model from a different lab (34% of AI natives and 37% of large orgs). By September it was about three-quarters (43% and 40%). Both lines rose mostly over the last two months. In Issue 02 we found that new models take share quickly, and this analysis shows the same is true from the buyer’s side: loyalty to a single model is weak, and loyalty to a single lab is getting weaker.

Certain workloads move to new models slowly

Share of accounts switching their top model for a given task, on average each month.
Share of accounts switching their top model for a given task, on average each month.

How often the main model changes depends a lot on what it’s used for. For each account and each task, we checked whether the top model for that task changed from one month to the next.

Coding and agent work churn the most. 53% of accounts changed their top model for general coding, 50% for multi-step planning, 49% for frontend work, and 45% for agent workflow execution. Coding is also where cross-lab moves are most common: 39% of accounts moved general coding to a different lab’s model.

Language and support tasks are the stickiest. Translation changed top model 24% of the time, summarization 27%, and conversation and customer support 29 to 30%. Only 11 to 15% of accounts moved those tasks to a different lab.

Sticky tasks are also more concentrated. For translation and summarization, the top model handles 81 to 83% of an account’s requests for that task. For general coding, it’s 64%.

One interpretation is that for translation, summarization and support, current models are already “good enough”, and switching means re-checking prompts and outputs that customers see for relatively little gain. For coding and agents, each new launch can noticeably change results so teams keep testing and moving.

Your model portfolio

Organizations are compiling model portfolios but each portfolio has a center of gravity. The typical org keeps about two-thirds of its traffic on one model, adds new models around the edges, and is not shy about replacing that center when a new flagship model arrives.

See which models are winning each task right now on /rankings.

Onwards,

Peter

Charts 1 to 3 cover organization accounts in the AI-native and large-org segments, October 2025 to September 2026, by request count. Models used = distinct models with at least one request that month. Top model = the model with the most requests in a month. Switch = a different top model than in the previous month, among orgs active in both months; a different lab = a different model author. The task chart covers all account types and classified traffic only: account-task pairs with at least 10 requests in both months, June→July, July→August and August→September 2026 pooled, for tasks with at least 50 pairs.