erabot.ai
ProductPricingCase studiesEvalBlogDocs
Book a demo

The default model is the expensive one: five products, one bug

TL;DR: We audited 15 open-source AI products for LLM cost waste. The single most repeated root cause was not a clever architectural mistake. It was that the default model was the expensive one, in five separate codebases, including one where the constant named DEFAULT_MINI_MODEL pointed at the most expensive model in the fleet. Nobody chose this. Defaults rot.

Why the default model matters more than the model picker

Most AI products let a user choose a model, and most users never do. So the default model is the production configuration for the large majority of requests. A default set carelessly, or set correctly two model generations ago, becomes the thing your bill is actually made of.

This is the kind of waste static analysis is good at finding, because it is in the code rather than the traffic. A dashboard shows you spent more this month. It cannot tell you that a constant three directories away names a flagship model for a job a small one would do.

Five of the fifteen products had it.

Five products, five expensive defaults

ProductWhat the default wasModeled saving
OpenHandsFrontend hardcoded Opus, the most expensive Claude model, as the default for every new user$5,040/mo
Lobe ChatDEFAULT_MINI_MODEL pointed at the most expensive model in the fleet$1,800/mo
LangflowAnthropic component silently defaulted to the flagship model$1,120/mo
CrewAIEmbedding defaults hardcoded the large embedding model in six or more places$250/mo
LibreChatTitle generation hardcoded a legacy model instead of a cheaper, better current one$60/mo

Those figures are modeled at stated call volumes, not observed bills. They scale with real traffic and are estimates until measured against a deployment's own usage.

How this happens

Lobe Chat's is the finding I would put on a slide. A constant called DEFAULT_MINI_MODEL resolved to the most expensive model in its fleet.

Read the name and you know exactly how it happened. Someone chose a genuinely small, cheap model, named the constant for its job, and moved on. Later the model string was updated to a newer generation. The new model carried "mini" in its name too, and it was not cheap. The constant kept its name. Every code review after that point saw DEFAULT_MINI_MODEL and read it as "the cheap one", because that is what it says.

No individual commit was wrong. The name stopped being true and nothing in the toolchain could notice.

Why defaults go wrong in exactly this way

Four mechanisms, all of which we saw:

A name outlives its meaning, as above. The variable says cheap and the value is not. Reviewers read names, not pricing tables.

The default is set where nobody shops. OpenHands set the model in the frontend, so the expensive choice lived in UI code rather than anywhere a cost-conscious engineer would look.

A component defaults to flagship to look good in a demo. Langflow's Anthropic component picked the strongest model when unspecified, which is the right call for a first-run experience and the wrong one for every request after it.

The expensive option is correct somewhere, so it spreads. CrewAI hardcoded the large embedding model in six or more places. Copy-paste carries the choice into contexts where nobody re-evaluated it.

What this says about where AI cost waste lives

Across all 15 products the audit raised 94 cost findings. Model and task mismatch was 3 of them, and the rest was structural: context re-sent uncached, calls made that did not need making, retrieval pulling more than the task consumed.

That looks like a contradiction with everything above, and it is worth resolving. The 3 figure counts call sites where the task was mismatched to the model. The defaults above are a different shape: one constant, read by thousands of requests. Few findings, large blast radius. Cheap to fix, which is the main reason to look for them first.

Fixing a default is a one-line change with no architectural risk. It is still a model swap, so it still needs verifying: a cheaper default that degrades output is not a saving. When our own agent proposed 31 model swaps across these products, the quality gate rejected 12 of them. We write about how that verification works in how to prove a model swap holds quality.

How to check your own defaults this afternoon

Three greps and a question:

  1. Find every place a model name is a string literal: your LLM client, and also UI code, deployment scripts, eval harnesses and test fixtures.
  2. For each one, ask what task it serves, then ask whether the smallest model you trust could do that task.
  3. Check whether any constant's name still describes its value. DEFAULT_MINI_MODEL is not an unusual mistake, it is the normal outcome of a model string being updated once.

Then verify before shipping, because a swap you have not tested is a candidate and not a decision.

If you would rather have this done by something that reads the whole repo, a free scan will find the call sites in a file, and the CLI runs the full audit in your own environment under your own model keys. Your source never reaches us.


Method: findings come from erabot's agentic audit of 15 open-source AI products, written up in full in we audited 15 AI products. Savings are modeled at stated call volumes and are estimates until measured against a deployment's real usage. Findings reflect the revision audited; maintainers may have changed these defaults since, and we filed findings upstream with patches attached.

EB
erabot.ai

Scan any codebase. See exactly what you spend on AI. Get one-click fixes.

No spam. Only launch updates and cost optimization tips.

Product

  • Product
  • CLI
  • Pricing
  • Changelog
  • Docs

Resources

  • Example audit
  • Blog
  • Case studies
  • Eval

Company

  • Security
  • Privacy
  • Terms
  • Support

© 2026 erabot.ai. All rights reserved.