Model announcements are full of numbers. Here is how to read the most common ones — and which ones deserve skepticism.

Parameters (“70B”, “405B”)

The number of learned values inside a model. More parameters generally mean more capacity, but training data, training method and architecture matter just as much. In a mixture-of-experts model, the announcement may give both a total parameter count and the smaller number of active parameters used per token — the active figure is closer to the real computing cost.

Tokens

Language models process text as tokens. In English, a token is on average about three-quarters of a word, so 1,000 tokens is roughly 750 words. Other languages and code can use more tokens for the same content.

Context window (“128K”, “1M”)

The maximum number of tokens the model can consider at once — your prompt, documents and its answer combined. A 1-million-token window can hold several long books. But fitting text in the window is not the same as using it well: quality on very long inputs varies, so test it on your own documents.

Price per million tokens

APIs usually charge separately for input tokens (what you send) and output tokens (what the model writes), and output is typically more expensive. Some providers discount repeated input through caching. Reasoning models may also bill for the hidden “thinking” tokens they generate.

Benchmark scores

Percentages on tests such as coding or science questions. Useful for rough comparison, but check who ran the test, under what settings, and whether results were independently reproduced.

Speed

Measured in tokens per second (how fast text streams) and time to first token (how long before the answer starts). For chat applications, both matter for perceived quality.

The numbers that matter most

For most real projects: quality on your own tasks, total cost per completed task, latency, and data-handling terms. The headline numbers are a starting point, not a verdict.

A worked example: reading a launch announcement

Imagine an announcement: “Our new 120B mixture-of-experts model (12B active) has a 256K context window, scores 88% on a reasoning benchmark, generates 150 tokens per second and costs $0.50 / $2 per million input/output tokens.”

  • 120B total, 12B active: large knowledge capacity, but compute per token similar to a 12B dense model.
  • 256K context: can read roughly a few hundred pages at once — but check how well it uses long inputs.
  • 88% on a benchmark: compare with other models on the same benchmark and settings; check whether results are independently verified.
  • 150 tokens per second: fast enough for real-time chat, depending on the provider’s hardware.
  • $0.50 / $2: input tokens are cheaper than output; estimate your monthly cost from expected usage.

Numbers that are often missing

  • Training data details — sources, size and cut-off date.
  • Results on independent evaluations rather than self-reported ones.
  • Reliability — variance across repeated runs.
  • Safety evaluation results and known limitations.
  • Energy use for training and inference.

Rules of thumb

  • About 750 English words ≈ 1,000 tokens.
  • A model needs roughly 2 bytes per parameter at 16-bit precision, so a 70B model needs about 140 GB of memory just for weights — less with quantization.
  • A doubling of context window does not mean a doubling of useful understanding.

Learn more in our AI terms everyone should know, the guide to GPU memory sizing for LLMs and the quick take on benchmark scores.

Bottom line

Read numbers relative to competitors and your needs, never in isolation.

What to watch next

  • Standardized reporting of model evaluations.
  • Independent benchmarking organizations.
  • Transparency requirements for general-purpose AI models.
  • Energy and efficiency metrics in announcements.

Key takeaways

  • Parameters indicate model size; active parameters matter for mixture-of-experts models.
  • Tokens measure text; about 750 words is roughly 1,000 tokens.
  • Context window size doesn’t guarantee good use of long inputs.
  • Benchmark scores need context and independent verification.
  • Price and speed determine practical usefulness for your workload.

Frequently asked questions

Is a bigger model always better?

Not necessarily. Smaller, well-trained models can outperform larger ones on many tasks and cost less.

What does “per million tokens” pricing mean?

You pay a set price for every million tokens processed, with separate rates for input and output.

How much memory does a model need?

Roughly 2 bytes per parameter at 16-bit precision, less with quantization, plus extra for context.