Methodology

This page describes how the data in Seminal AI is gathered, normalised and maintained, and what its limits are.

Where the figures come from

Every price, context window and capability figure is taken from the vendor's own official pricing page or API documentation. We do not use press coverage, aggregator sites, or secondhand summaries as a primary source. If a vendor does not document it publicly, it does not appear here.

Each model entry records two things alongside the data: the source URL the figure was read from, and the date it was last checked. Both are shown on the entry. You can follow the link and verify the figure against the vendor at any time.

How prices are normalised

Vendors publish pricing in inconsistent units — per thousand tokens, per million tokens, per character, per request, in different currencies. We normalise everything to USD per million tokens so that entries can be compared side by side. Where a vendor prices in another currency, the entry notes it.

A dash in a price field means the vendor does not publish a figure for that item. It does not mean zero, and it does not mean free. Treat a dash as "unknown, ask the vendor."

Open-weight models show no per-token price. This is not missing data. A model you download and run yourself has no per-token cost — it has hardware, operational and engineering costs that depend entirely on your deployment. Where a hosted provider serves an open-weight model, that provider's price appears under the provider, not the model.

Blended cost

Comparing models on input price alone is misleading, because output tokens are usually priced several times higher and real workloads consume both. We publish a blended cost computed at a 3:1 input-to-output ratio — three input tokens for every output token.

That ratio is an approximation of common production usage: prompts, context and retrieved documents typically outweigh generated text. It is not correct for every workload. A summarisation pipeline skews further toward input; a long-form generation task skews toward output. The blended figure is a fairer single number for ranking than input price alone, but if your ratio is materially different, compute your own from the two published prices.

Verification and cross-checking

Entries are re-verified periodically. Each entry is also cross-checked by a second independent pass before publication, reading the vendor source afresh rather than confirming the first reading.

Even so, this data can lag. Vendors change prices, deprecate models and ship new ones without notice, and a checked date of six weeks ago is a real gap. Confirm current pricing directly with the vendor before committing to a contract or building a budget on it. This manual is a starting point for a decision, not a substitute for the vendor's terms.

What we do not do

We do not run our own benchmarks. We do not have the infrastructure to run evaluations at a standard we would be willing to publish, and we would rather publish nothing than publish numbers we cannot defend. Where benchmark results appear, they are attributed to whoever produced them.

We do not accept payment for inclusion or ranking. No vendor can buy an entry, an improved position, or a more favourable assessment. Rankings follow the stated criteria on each page and nothing else.

Corrections

Corrections are welcome and are treated as bugs. If a figure here disagrees with the vendor, tell us: info@seminal.ai. Include the entry and, if you have it, the vendor URL showing the correct value.