A useful shortlist.
Honest about its limits.
Here’s what the tiers mean, where the numbers come from and what we still need to test.
Back to rankingsCapability and task fit are different.
S+ and S sit under the frontier/SOTA heading. A and B are workhorses. C covers lightweight models. These are editorial capability classes for this guide, not scores from trade-business benchmarks.
A lightweight model can still be the sensible choice for extracting invoice fields or drafting a short update. Choosing a task filters the shortlist; it does not turn an untested opinion into a measured ranking.
Facts from sources. Judgement kept separate.
Model identities, context and base API prices come from a dated public OpenRouter catalogue snapshot, checked against provider information. Prices are US dollars per million input or output tokens. Long-context, cache, provider-route and other conditions can change the actual bill.
OpenRouter’s usage leaderboard measures tokens processed on its platform. It does not establish accuracy, trade suitability, user numbers or market-wide adoption. A subscription app can also expose different tools and limits from its underlying model.
What a real trade test needs.
The proposed pilot covers quote scopes, job reports, marketing and bookkeeping preparation across electrician, panel-repair and automatic-door examples. Twelve cases, three models and two runs produce 72 outputs for human review.
- Fix the source facts, required fields and unacceptable errors before testing.
- Keep tools and inputs comparable. Record versions, settings, failures, time and cost.
- Have the right person review factual fidelity, completeness and editing effort.
- Publish examples, failure counts and ties. Keep native-app tests separate.
Invented prices, wrong job identities and unsupported commitments block a recommendation. Small samples do not justify fine-grained scores. No paid evaluations have been run for this guide.
What Web Blend gets out of this.
We help trade businesses organise information and connect the software they already use. This resource is designed to help you choose well, and to make it easy to ask us for implementation help. An existing native feature may be all you need.
Talk through my workflow