Terry Djony publishes a daily-updated model shopping table using Artificial Analysis prices and scores: Reveneau

Why it mattersA team picking a model for a workflow usually reads a benchmark once and a price list once, and the two are never together. This puts both on one screen and updates them every day.

Picking the model to run a new agent on takes reading two lists that are updated separately: a benchmark and a price sheet. Terry Djony's site puts both on one screen and re-fetches the numbers every night.

Djony posted the site to Hacker News this morning and by mid-afternoon Pacific it had 131 points. The page charts every model in the Artificial Analysis Intelligence Index against its blended API price, on a log scale, and highlights the "value frontier": the models where nothing cheaper is also smarter.

What it shows

The main chart is a scatter of price against Intelligence Index score. The blended price uses a three-to-one input-to-output token mix and is drawn per one million tokens. Points on the frontier are marked with an accent colour and connected with a line, so a reader can trace the curve of best-available quality by budget.

Below the chart is a lookup table Djony calls the budget picker. Each row is a price band. The pick is the highest-scoring model that fits, and the runner-up is the next-best model that also fits. Below that is a raw capability bar chart, sorted by Intelligence Index alone.

Two smaller sections are the useful part for a team using the site over time. A "What changed" panel diffs consecutive daily fetches and lists new models, removed models, and models whose score or price changed. An "All figures" table is sortable by column, so a reader can rank by output-token price or by score directly.

Where the numbers come from

The site says the data source is the free data API from Artificial Analysis, fetched daily by a GitHub Actions cron. Artificial Analysis publishes its own model index; Djony's site does not run its own benchmark, so the scores are theirs and the site is a lookup on top.

The code and data are open at github.com/terryds/bestvaluemodel. Repository star count was five at the time of writing, four hours after creation, so the repo itself does not carry the traction. The Hacker News thread does.

The one thing to check before trusting a row

Blended price hides output-heavy workloads. A model that is cheap on a three-to-one input-to-output mix can be expensive on a workload that produces long answers per short prompt, because output tokens usually cost several times what input tokens cost. The site says the ratio out loud; a team should re-rank with its own ratio before committing to a switch.

The "value frontier" definition is worth reading in the "How to read this" section. Djony sorts models by price ascending and keeps every model that scores higher than everything cheaper. Ties on price go to the higher score. Ties on score go to the cheaper model. That is a strict "no cheaper model is also smarter" rule, which is the right one for a shopping question, and worth noting because a model just below the line can still be a fine choice for a specific job.

The consequence

A team picking a coding, chat or agent model this week can now open one page and see the current top pick for its budget, refreshed against Artificial Analysis's own daily update. Vendors cut prices often enough that a decision made in September will need to be checked again in October.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.