What you can ask

The consultant recognises eight kinds of question. You don't need to pick one — phrase the question naturally and it routes itself.

Recommendation

“Best model for agentic coding.” Ranks models on a composite of the benchmarks that matter for that task and returns a shortlist with a named pick.

Superlative with a constraint

“Cheapest model with vision and 200K context.” Optimises one axis — cheapest, fastest, longest-context, most capable — among models that clear your constraints. When you also name a task, candidates must pass a competence bar first, so the cheapest answer is never a model that can't do the job.

Head-to-head comparison

“Compare Claude Opus 5 and GPT-5.2.” Puts two to four named models side by side on price, intelligence, coding, agentic, speed and context, then says which to pick under which conditions.

Model profile

“Tell me about Claude Sonnet 5.” Full specification, percentile ranks across the catalogue, providers, licence, and a cheaper near-equivalent if one exists.

Price-performance frontier

“Where is the price-performance frontier?” Plots price against the intelligence index and identifies the models where nothing cheaper is smarter, plus the best option in each price tier.

Cost projection

“What does 2 million requests a month cost?” Projects monthly spend across candidate models at a stated volume and token mix, and flags where prompt caching would change the maths.

Landscape

“Who leads the model landscape?” Aggregates the catalogue by maker — model counts, strongest model, median price, open-weights share.

Filtered listing

“Which models support audio input?” Returns everything matching a constraint, ordered by capability then price.