About

The problem this solves

Comparing model costs should be simple arithmetic and is not. Providers publish prices in different places and different units — per million tokens, per minute of audio, per image, per second of video. Headline figures come with conditions that materially change the bill: promotional windows with expiry dates, thresholds that reprice an entire request once a prompt crosses a line, retention requirements that rule out certain deployments entirely.

And prices move. During 2026 one budget tier fell 80% in a single day and a flagship dropped over 20% six weeks later. Any table quoting launch-day figures can be wrong by a factor of five, which describes a surprising amount of what is published about model pricing.

What we do differently

• Dates on everything. Every price shows when it was last checked against a primary source. A comparison site without a verification date is asking you to trust a snapshot of unknown age.
• Conditions recorded, not hidden. Promotional expiry dates, long-context repricing thresholds, and retention requirements appear on the model page rather than in a footnote.
• Native units kept. Transcription bills per minute and video per second. Forcing those into a token-equivalent produces a tidy table and a misleading one.
• Omission over guessing. If we cannot verify a price from the provider, the model is left out. Unreleased models with rumoured specifications are not listed at all.
• No commercial relationships. No affiliate links, no sponsored placements, no paid rankings.

The full process is written up in the methodology.

What 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 — spanning language, speech, image, and video categories, plus head-to-head comparisons and a cost calculator for modelling your own token volumes.

What we deliberately avoid

We do not rank models primarily by benchmark position. Leadership rotates every few weeks, and single-shot scores systematically overestimate how reliably a model performs across varied real attempts. Where we publish benchmark numbers we label the source and say plainly that twenty examples from your own domain will tell you more than any public leaderboard.

We also avoid "best model for X" verdicts that ignore cost. The most capable model available currently leads the main intelligence index by roughly one point over a competitor costing less than half as much per finished task — a fact that matters more to most teams than the ranking does.

Built as a static site, on purpose

No accounts, no logins, no paywall, and no backend. Pages are pre-rendered and served from a CDN, which keeps them fast and keeps the amount of data we could collect about you close to zero. See the privacy policy for what that means in practice.

Corrections welcome

A stale price is the worst thing a comparison site can publish, so corrections take priority over everything else. If a figure here is wrong, send the provider link via the contact page.