The Model Register
Nineteen language models seen two ways: how each is built, and what it means to depend on it. Layer stacks, KV cache and GPU memory for CTOs; prices, licenses, jurisdiction and vendor concentration for boards.
This page shows nineteen. Our full analysis covers more than 100 models, and the complete register is shared with clients during the consulting phase of an engagement.
Will it run on your hardware? See the Compute Register case study
Model families
Nineteen models grouped four ways. Select a ring to zoom in, the center to zoom out, and a model on the outer ring to open its card.
How each model is built
Each bar shows a model’s decoder layers in order of type, drawn to the same scale, so a longer bar is a deeper network. Closed models do not publish their architecture. Select a name to open its card.
Register
Every card pairs the engineering facts with the questions a buyer has to answer: who controls the logs, who can change the model under you, and where the provider is based.
A sample of the full register. We have analyzed more than 100 models the same way. The complete register is shared with clients during the consulting phase of an engagement, under a mutual non-disclosure agreement.
Compare two models
Highlighted rows differ. The bars underneath line the two decoder stacks up at the same scale.
What will it cost?
List prices tell you little until they meet your own prompts. Token Bench, our cost calculator, counts your text with each model’s real tokenizer and prices it across these models at your volume, including caching, batch discounts, reasoning tokens, and routing between a cheap and a strong model.
- Your prompt, not an average. Paste real text and see token counts per model, computed in your browser.
- Your language. The same text in an Indian language can cost several times as much in tokens. Token Bench measures this across fourteen Indian languages.
- Your volume. Monthly cost by model, with announced price changes applied on their dates.
GPU memory planner
For open-weight models you host yourself. Weights scale with total parameters and precision; the KV cache scales with context length and concurrent requests.
Vendor and jurisdiction mix
Tick the models your organization runs and give each a share of traffic. The check shows how much depends on one provider or one country, and whether you have a path to self-hosting.
Method and sources
Prices
List prices per million tokens from each provider’s pricing page, checked in October and November 2026, on the same basis as Token Bench: standard rates, DeepSeek at off-peak. Introductory rates are marked with their end date. For the cost of your own prompts, use Token Bench.
Architecture
Parameter counts, layer mix, attention design, and KV cache per token come from each model’s published configuration. KV figures are for BF16 caches and count only layers that keep a cache; recurrent layers hold a small fixed state.
Record and change control
Derived from how a model is delivered. With open weights you can keep every input and output yourself and pin the version. Through a closed API, logs pass through the provider and the model can change under you, which in the Vector & Value framework resets the delegation stage.
Matrix fit
Our editorial view of which Vector & Value Matrix zones each model suits today. It will be replaced by ReasonIQ audit results as they are completed.
This register is for information. It is not a recommendation to buy or use any model, and license terms should be read in full before commercial use.
Prices last checked 4 November 2026. Model names and trademarks belong to their owners.
Want the full register of 100+ models?
The complete analysis, with the same engineering and business view for every model, is shared with clients during a consulting engagement, under a mutual non-disclosure agreement.