Methodology
A comparison site is only worth reading if you can trust the numbers. Here is exactly where ours come from and what we do when sources disagree.
Pricing checked against provider documentation on . How we verify
Primary sources first
Prices come from the provider's own pricing page or API documentation — not from press coverage, not from other aggregators. Where a provider publishes both a launch price and a subsequent revision, we record the current rate and keep the launch price in the caveat field so the change is visible rather than silently overwritten.
We currently track 35 models across 13 providers: Anthropic, AssemblyAI, Black Forest Labs, DeepSeek, Deepgram, ElevenLabs, Google, Ideogram, Kuaishou, Moonshot AI, OpenAI, Runway, Z.ai.
Normalization, and its limits
Language models are normalized to US dollars per million input and output tokens, which makes them directly comparable. Other categories are not, and we do not pretend otherwise: transcription bills per minute of audio, synthesis per character, image models per generation, and video per second of output. Forcing those into a token-equivalent produces a tidy table and a misleading one, so each model keeps its native unit.
For ranking purposes only, we compute a blended cost assuming a 3:1 input-to-output ratio, which is representative of chat and agent workloads. That figure drives the "cheapest first" sort and nothing else.
What the caveats field is for
Headline prices increasingly come with conditions, and 30 of the models we track carry at least one. Promotional rates with expiry dates, long-context thresholds that reprice an entire request, data retention requirements that rule out zero-retention deployments — these change the real cost more than a few cents per million tokens does. Where a condition exists, it appears on the model page rather than in a footnote.
Benchmarks are treated as directional
We list benchmark scores where a provider or an independent index has published them, and we label which. We do not compute our own scores, and we do not rank models primarily by benchmark position — leadership rotates every few weeks and single-shot metrics systematically overestimate how reliably a model performs across varied real attempts.
The recommendation we would give in every case: run twenty examples from your own domain. That beats any public leaderboard for deciding what to ship.
When sources conflict
It happens, particularly around launches and unreleased models. Our rules: the provider's own documentation wins over its blog post; a dated pricing page wins over an undated one; and if we cannot resolve a conflict from a primary source, we leave the model out rather than average the guesses. Unreleased models with rumoured specifications are not listed at all, however widely those figures circulate.
No commercial relationships
No affiliate links, no sponsored placements, no paid rankings. We are not affiliated with any provider listed here. Outbound links to provider documentation are marked nofollow.
Corrections
If a number here is wrong, we want to know — a stale price is the worst thing a comparison site can publish. Report it via the contact page with a link to the source and we will correct it.
Frequently asked
How often is pricing updated?
Every model is re-checked against the provider's own pricing or documentation page, and the verification date is shown on each table and model page. Provider prices moved several times in 2026 — including an 80% cut to one budget tier — so a date stamp matters more than a claim of being 'always current'.
Why do some prices differ from other comparison sites?
Usually because they quote launch prices rather than current ones. When a provider cuts a price mid-cycle, aggregators that scraped the announcement post keep showing the old figure. We record the current rate and note the launch price in the caveat field.
Do you earn commission from providers?
No. There are no affiliate links, sponsored placements, or paid rankings on this site, and we are not affiliated with any provider listed. Links to provider documentation carry nofollow.
Why is a model I use missing?
Most likely because we could not verify its current price from a primary source, or it has been superseded. We would rather omit a model than publish a number we cannot stand behind. If something is missing, tell us.