{"id":2441,"date":"2026-08-25T17:50:01","date_gmt":"2026-08-25T17:50:01","guid":{"rendered":"https:\/\/www.danielpradilla.info\/blog\/?p=2441"},"modified":"2026-08-25T17:50:04","modified_gmt":"2026-08-25T17:50:04","slug":"minimum-viable-model","status":"publish","type":"post","link":"https:\/\/www.danielpradilla.info\/blog\/minimum-viable-model\/","title":{"rendered":"Minimum Viable Model"},"content":{"rendered":"

A lot of the discussion around AI model routing starts from a common idea: don\u2019t use the most powerful model if a cheaper one can do the job<\/strong>.<\/span><\/p>\n

Microsoft Research has explored this multiple times. In Hybrid LLM<\/span><\/a><\/em>, BEST-Route<\/a><\/em>, One Head, Many Models<\/a><\/em> and, more recently in 2026, Switchcraft<\/span><\/a><\/em>, the objective is broadly the same: route each request to the lowest-cost model that can still meet the required quality or correctness threshold.<\/span><\/p>\n

You can see the same broad idea in systems such as GitHub Copilot\u2019s automatic model selection.<\/p>\n

It is something like a Minimum Viable Model<\/strong> principle.<\/p>\n

Recently I ran into a paper, The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs<\/span><\/a><\/em>, that made me realize that agentic<\/span> workflows make this problem harder.<\/span><\/p>\n

Small differences in per-step reliability can translate into very large differences in the probability of successfully completing a long sequence of actions.<\/span><\/p>\n

If a model is 99% reliable at each step, then over 100 independent steps:<\/p>\n

0.99\u00b9\u2070\u2070 ~ 37%<\/strong><\/p>\n

At 99.9%:<\/p>\n

0.999\u00b9\u2070\u2070 ~ 90%<\/strong><\/p>\n

And real agents can be worse than this simplified example, because errors are not necessarily independent: a mistake can enter the context and influence later decisions.<\/p>\n

So the routing question changes from:<\/p>\n

What is the cheapest model that can answer this request correctly?<\/strong><\/p>\n

to:<\/p>\n

What is the cheapest model that can complete the expected trajectory with an acceptable probability of success?<\/strong><\/p>\n

That is a different optimization problem.<\/p>\n

The cheapest model per token is not necessarily the cheapest model per answer. And the cheapest model per answer is not necessarily the cheapest model per completed task.<\/p>\n

A weaker model may generate more reasoning tokens, require more retries, make more tool-call mistakes, accumulate bad context, or eventually need to be replaced by a stronger model.<\/span><\/p>\n

In coding agents like copilot Model switching during a long-running session is not free, because rerouting can mean processing a lot of context again.<\/span> GitHub therefore performs automatic model selection at “natural boundaries” rather than continuously changing models during an active task: GitHub Copilot automatic model selection<\/a>.<\/span><\/p>\n

So a routing decision made near the beginning of a task can have consequences much later.<\/p>\n

Microsoft Research\u2019s Switchcraft<\/span><\/a><\/em> describes routing as selecting the lowest-cost model subject to correctness<\/strong>.<\/p>\n

But for agentic<\/span> systems, that would translate to:<\/p>\n

Selecting the lowest-cost model subject to trajectory-level correctness.<\/strong><\/p>\n

A model that is only slightly less reliable may look almost as good on a benchmark or on a single question.<\/span> But over a long chain of actions, those small differences compound. This is also the paper\u2019s argument for continued scaling: \u201cscaling model size leads to large improvements in the number of turns a model can execute successfully.\u201d<\/strong><\/span><\/p>\n

So perhaps the real routing question is not simply which model can do this task?<\/em> but which is the cheapest model I trust to get all the way to the end of the job?<\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"

A lot of the discussion around AI model routing starts from a common idea: don\u2019t use the most powerful model if a cheaper one can do the job. Microsoft Research has explored this multiple times. In Hybrid LLM, BEST-Route, One Head, Many Models and, more recently in 2026, Switchcraft, the objective is broadly the same:… Continue reading Minimum Viable Model<\/span><\/a><\/p>\n","protected":false},"author":2,"featured_media":2442,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_feature_clip_id":0,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_post_was_ever_published":false},"categories":[174],"tags":[],"class_list":["post-2441","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-bestof","entry"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t