Methodology
A lifecycle dataset is only worth as much as its sourcing. Here is exactly where the numbers and dates come from, how they are checked, and what this dataset deliberately does not contain.
Sources
Every record is compiled from the provider's own published material: the model list, the pricing page, and the deprecation or migration notice. Those pages are the authority — not a third-party summary, and not another comparison site.
Each model stores the URLs it was built from, and every page on this site links them. If our record and the provider's page ever disagree, the provider's page is right and ours is a bug.
First-party versus aggregator
Some providers in this dataset are not model makers but routers and cloud platforms — OpenRouter, Amazon Bedrock, Azure AI Foundry, Vercel AI Gateway and the like. Their prices are real, but they are the cost of routing through that platform, not the maker's official tariff, and their lifecycle can differ from the maker's.
So every record carries a source type. Of the 354 models tracked, 310 come from a first-party source and 44 from an aggregator listing. Aggregator records are labelled as such wherever they appear, and are never presented as an official price.
Verification
Every model carries a verification timestamp — the moment a human or a collector last confirmed it against the source — and a verified flag. Both are shown on the model page. A record we could not confirm is marked unverified rather than quietly published as fact.
Where a provider announces a deprecation without committing to a shutdown date, we record the deprecation date and leave the shutdown date empty. We do not estimate, interpolate or infer dates. An empty field means the provider has not said, which is itself useful information.
How the pipeline runs
- Collect. A collector per provider produces that provider's model list in a fixed contract. Today every collector reads a reviewed dataset compiled from the sources above; automated fetching is being rolled out provider by provider, starting with the ones that publish a machine-readable model API (Anthropic, OpenAI, Google).
- Validate. The assembled dataset is checked against a schema and a set of plausibility rules before anything is stored. This is a hard gate: a run that fails validation is aborted and the previous data stays in place. Unexpected structure never passes through silently.
- Detect changes. Import is idempotent. A price snapshot is written only when a price actually changed, and a status event only on a real transition — so the history records changes, not the fact that a job ran.
- Publish. The dashboard reads the stored data directly. This marketing site is static, so its lifecycle pages reflect the dataset as of the last build — every page that shows a time-sensitive figure states the snapshot date it was built from.
What we do not do
- No quality benchmarks or evals. We do not rank models by how good they are. Lifecycle and price are verifiable facts; "best model" is not.
- No paid placement. No provider can pay to appear, to rank higher, or to have a status softened. There is no advertising on this site.
- No estimated dates. If a provider has not announced a shutdown date, the field stays empty. A plausible-looking guess in a date field is worse than a blank one.
- No silent corrections. Price and status changes are stored as dated events, so a change is visible as a change rather than replacing history.
- No scraping behind logins or paywalls. Only public provider documentation.
Known limits
Being straight about this matters more than looking complete:
- Refresh is scheduled, not real-time. A provider announcement made an hour ago may not be reflected yet — which is why every record shows when it was last verified.
- Coverage is 20 providers, chosen by usage rather than by trying to list every endpoint that exists. Missing one you rely on is a reasonable thing to tell us about.
- Lifecycle is tracked per provider. The same model can be active on one platform and retired on another; we record each separately rather than merging them into one misleading row.
- Prices are list prices in USD per million tokens. Volume discounts, committed-use agreements, batch and cache tiers are not modelled yet.
Corrections
If a record is wrong, the fastest fix is the provider link on the model page plus a note about what it should say. Corrections to lifecycle dates take priority over everything else on our list — a wrong shutdown date is the one error in this dataset that can actually break something.