Methodology

Every number on this site is calculated client-side from a registry of model specs and prices verified against primary sources. This page explains how — including where results are exact and where they are estimates.

Token counts

OpenAI models — exact. We run the o200k and cl100k BPE encodings in your browser via the gpt-tokenizer library. We verify on every release that it produces byte-identical token IDs to OpenAI's official tiktoken library across a multilingual test corpus (English prose, code, JSON, German, Japanese, Russian, emoji and whitespace edge cases). One caveat: OpenAI has not published the encoding used by the GPT-5.4/5.6 families; we use o200k by lineage, which matches all published GPT-5 models.

Claude and Gemini — labelled estimates. Neither provider publishes a client-side tokenizer, so we estimate from their official guidance: Anthropic states a Claude token is roughly 3.5 English characters; Google states a Gemini token is about 4 characters. Real counts vary with content type — code and non-English text tokenize less efficiently — and Anthropic notes its Claude 4.7+ tokenizers produce around 30% more tokens than earlier generations for the same text. Wherever a count is an estimate, the tool says so. For billing-critical work, use the provider's official count-tokens API.

Other models — generic estimate. Models without a published browser tokenizer or ratio use the common ~4 characters per token approximation.

VRAM requirements

Total VRAM = model weights + KV cache + framework overhead.

Weights. Parameters × bytes-per-parameter. Rather than theoretical bit-widths, our multipliers are measured from actual GGUF quantised files published on Hugging Face (popular K-quant releases, exact byte sizes from the file listings):

Measured this way, predictions land within about ±1.2% of real file sizes. K-quants sit above their nominal bit-width because important tensors (embeddings, attention) are kept at higher precision. Mixture-of-experts models follow the same rule using theirtotal parameter count. Models that ship pre-quantised (GPT-OSS in MXFP4, Mistral's FP8 releases) are calculated at their native format.

KV cache. We model each architecture as published in its config.json rather than assuming classic dense attention:

• Dense GQA (Llama, Phi, most pre-2026 models): 2 × layers × kv_heads × head_dim × context × 2 bytes (FP16 cache). This matches llama.cpp's reported KV size exactly (e.g. 1024 MiB for Llama 3.1 8B at 8K context).
• Hybrid attention (Gemma 4, Qwen 3.5/3.6, GPT-OSS): only full-attention layers cache the whole context; sliding-window layers cap at their window and linear-attention layers hold constant-size state. A flat formula would overestimate these models 4–6×.
• Latent compression / MLA (DeepSeek V4, Mistral Large 3, GLM-5.2): layers × (latent rank + RoPE dim) × context × 2 bytes — the design that makes million-token contexts feasible.

Overhead. We add 10% plus a flat 1GB for the inference framework's compute buffers, matching typical llama.cpp behaviour. For Apple Silicon we count 75% of unified memory as GPU-available (macOS's default Metal limit; advanced users can raise it with the iogpu.wired_limit_mb sysctl).

Treat results as realistic planning figures, not guarantees — actual usage varies slightly by runtime, driver and settings.

API costs

Prices are per-million-token list prices taken directly from each provider's official pricing page (OpenAI, Anthropic, Google, DeepSeek, Mistral) — never from third-party comparison sites. Where pricing has known caveats (introductory rates, prompt-length tiers, cache storage fees), the tool shows a note alongside the price.

Update cadence

The model registry records the date each entry was last verified, and our build fails if any entry goes more than 60 days unchecked — staleness is treated as a bug. We re-verify prices monthly and add newly released models within days. Every page shows the registry's last-updated date (2026-08-03 as of this build).

Privacy

All tools run entirely in your browser. Text you paste and inputs you choose are never uploaded, logged or stored — see theprivacy page.

Last updated 2026-08-03.