What we build in the OpenAI API
Strong on classification and extraction, with a good range of model sizes for cost control. One of two providers we build with, chosen per task on measured accuracy rather than preference.
What it is, and where it sits
OpenAI's API is the other AI provider we use. In practice the choice between it and Claude is made per task by measurement, not by allegiance, and we build so the provider can be swapped without rewriting the surrounding workflow.
Where it tends to fit: classification and extraction at low cost on the smaller models, function-calling patterns, and workloads where the model tier can be tuned tightly against a measured accuracy floor.
What we build in it
- Inbound message classification at volume, on a small, cheap model
- Field extraction from semi-structured text into typed output
- Lead qualification from free-text enquiry into structured CRM fields
- Content summarisation for internal digests
- Semantic search over your own documents using embeddings
- Transcription workflows feeding downstream automation
How it is priced
Per token, input and output, varying substantially by model. Paid directly to OpenAI; we take no margin.
The meaningful lever is model selection. The gap between the smallest and largest models is large enough that running a classification task on an oversized model is one of the more common sources of unnecessary spend. We benchmark the small model first and only move up if measured accuracy requires it.
Embeddings for semantic search are cheap enough to be a rounding error for most SMB corpora.
When it is the right choice
- High-volume classification where a small model meets the accuracy bar cheaply
- Structured extraction using function calling
- Semantic search over a document corpus
- Transcription feeding a downstream workflow
- Where measurement on your data shows it outperforms the alternative — which is the only reason that really counts
Limits worth knowing before you commit
- Non-deterministic, like any model. Validate downstream; never trust the shape of the output without checking it.
- Model deprecations happen. Pin versions and plan for migration; this is a fast-moving market.
- Rate limits vary by tier and matter for bulk processing.
- Not a verification step for itself. Same rule as every provider.
Under the commercial and enterprise terms of both providers we use, your data is not used to train their models. If your compliance position requires that nothing leaves your infrastructure at all, we can work with self-hosted open-weight models instead, at some cost in capability. That gets decided during scoping, not afterwards.
We hold no reseller agreement with this vendor, take no affiliate commission, and do not resell licences. You buy your own subscription directly. It is the only arrangement under which our recommendation means anything.
Want the cost model run on your numbers?
Tell us your workflows and volumes and we will show you the three-year cost across every option.