All Embedding Models at a Glance

Model	Provider	Dims	Max Tokens	Price/1M tokens
`text-embedding-3-large`	OpenAI	3072	8191	$0.13
`text-embedding-3-small`	OpenAI	1536	8191	$0.02
`text-embedding-ada-002`	OpenAI	1536	8191	$0.10
`gemini-embedding-2`	Google	768	2048	Free*
`gemini-embedding-001`	Google	768	2048	Free*

*Google Gemini embeddings are free under API rate limits. Paid plans available for higher throughput.

Cost Analysis: Monthly Spend by Volume

Volume	ada-002	3-small	3-large	Gemini
1M tokens	$0.10	$0.02	$0.13	$0.00
10M tokens	$1.00	$0.20	$1.30	$0.00
100M tokens	$10.00	$2.00	$13.00	$0.00
1B tokens	$100.00	$20.00	$130.00	$0.00

The Catch: Vectors Are Not Compatible

You can't just swap text-embedding-3-small for gemini-embedding-2 and query your existing database. Even models with the same dimensions produce vectors in completely different mathematical spaces.

Cross-querying incompatible vectors returns garbage results:

Query Model	Index Model	Recall@10
3-small	ada-002	0.0% (same vendor, same dims!)
3-large	gemini-2	0.0%
3-small	3-small	100% (same model)

How to Switch Models Without Re-embedding

Schift uses projection matrices to map vectors between model spaces, achieving 92-104% retrieval recovery (yes, sometimes projections improve results):

Migration Path	Recall@10 Recovery	Verdict
ada-002 → 3-small	99.2%	SAFE
ada-002 → gemini-2	97.8%	SAFE
3-small → gemini-2	99.7%	SAFE
3-large → gemini-2	98.1%	SAFE
3-small → 3-large	104.2%	SAFE

Decision Matrix: Which Model Should You Use?

Choose Gemini if:

Cost matters — free under rate limits
768 dimensions are sufficient (most RAG use cases)
You want multi-provider redundancy
You're starting a new project

Choose OpenAI 3-small if:

You need 1536 dimensions for precision
You're already on OpenAI and cost is acceptable ($0.02/1M tokens is cheap)
Longer context windows matter (8191 tokens vs 2048)

Choose OpenAI 3-large if:

Maximum retrieval precision is critical
Your domain is highly specialized (legal, medical, scientific)
Budget is not a constraint

Avoid ada-002:

Legacy model — 5x more expensive than 3-small with no quality advantage
If you're still on ada-002, migrating to 3-small or Gemini saves money immediately

Migrate Between Any Models

pip install schift

from schift import Schift
s = Schift(api_key="sk-...")

# Test any migration path
report = s.bench.run(
    source="openai/text-embedding-ada-002",
    target="google/gemini-embedding-2"
)
print(f"Recovery: {report.recall_at_10}")

# If satisfied, migrate
s.upgrade(db="my_db", to="google/gemini-embedding-2")