MetaGPT ships you a config that defaults to gpt-4-turbo — for a tool that runs a whole AI software team
Third in the series. Swarm had a legacy model on a trivial call; gpt-researcher was well-architected. MetaGPT (40k+ stars — "one line of requirement → a simulated software company") shows the third place cost hides: not in the code, but in the example config you're told to copy.
The finding
MetaGPT's onboarding is: copy config/config2.example.yaml to config2.yaml, drop in your API key, go. Here's what that example config points you at:
# config/config2.example.yaml — the config you copy to start
llm:
model: "gpt-4-turbo" # or gpt-3.5-turbo
And its embeddings default (metagpt/learn/text_to_embedding.py, tools/openai_text_to_embedding.py):
async def text_to_embedding(text, model="text-embedding-ada-002", ...):
Both are legacy models that newer options beat on price and quality:
| In the config | Price (per 1M) | The modern replacement | Price |
|---|---|---|---|
gpt-4-turbo | $10 / $30 | gpt-4o (better model) | $2.50 / $10 → ~75% / 67% less |
gpt-3.5-turbo (the "or") | $0.50 / $1.50 | gpt-4o-mini (better) | $0.15 / $0.60 → cheaper and smarter |
text-embedding-ada-002 | $0.10 | text-embedding-3-small (better) | $0.02 → 5× less |
Nothing here is exotic — gpt-4o is a strict upgrade over gpt-4-turbo (OpenAI's own positioning), and 3-small beats ada-002 on both axes. The config just hasn't kept up.
Why it matters more here than usual
MetaGPT's whole premise is that one task spins up a team — product manager, architect, engineer, QA — each an agent, each making multiple LLM calls, iterating. So the model you pick in that config isn't one call; it's the multiplier on every role, every step, every run. A default that's 4× the price of a better model doesn't add 4× to one call — it adds it to the entire simulated company.
That's the pattern: the per-call mistake is small; the framework-wide default is where it compounds.
The fix
In your config2.yaml:
llm:
model: "gpt-4o" # strict upgrade over gpt-4-turbo, ~70% cheaper
# embeddings: text-embedding-3-small # 5× cheaper than ada-002, better
For the lighter roles (short reasoning, routing, structured fills), gpt-4o-mini is the bigger win again. As always: a model swap is a candidate — run it on your own tasks and confirm the output quality holds before you standardize on it. For a general upgrade like turbo→4o it almost always does; the point of shadow-testing is the cases where it's less obvious.
The actual lesson
Your real model decision lives in a config default, and defaults go stale. Most users never change the example config — they copy it and run. Models get cheaper and better every few months; the YAML in the repo doesn't. So the single highest-leverage thing a framework maintainer can do for their users' bills is keep the example config current — and the highest-leverage thing you can do is check what your framework actually defaults to, because it's probably a model generation or two behind.
Scan your own stack (free)
- Claude Code:
claude plugin marketplace add erabot-ai/erabot-plugin→/erabot:scan. It resolves the model even when it's set via config or env — which, as MetaGPT shows, is exactly where the decision actually lives. - Want the swaps verified on your real traffic + apply-ready diffs? → erabot.ai.
How erabot stays honest: assumptions are marked, prices are current list prices you can check, a swap is a candidate until shadow-tested, and we don't inflate a stale default into a scandal — it's a one-line fix, and we say so.