Claude (Anthropic)

What it is

Claude is a family of models from Anthropic, sold two ways: as the claude.ai assistant app and as an API that other products build on. Anthropic groups the family into lines: Haiku, Sonnet, Opus and Fable. Each line trades speed and price against reasoning depth.

How it works

You send text (and images) and get text back. On the API you pick a model ID and an effort level. Current models decide for themselves how much to “think” before answering, steered by that effort setting. Tool use lets the model call functions you define, which is how it reads files, searches, or acts inside other systems. The same models are offered through Amazon Bedrock, Google Cloud and Microsoft Foundry, so a company can buy through a cloud contract it already has.

Specs and pricing (checked 2026-10-17). All current models have a 1M-token context window and up to 128K output tokens. API list prices per million input and output tokens run from $0.10 and $0.50 (Haiku 5.5) to $10 and $50 (Fable 5.1), with batch processing at half price. Anthropic notes that the tokenizer used from Claude 4.7 onward produces roughly 30% more tokens for the same text, so compare cost per task, not just price per token.

Benefits

• Long inputs. A 1M-token window holds large document sets in one request (see context window).
• Clear tiering. Price and speed steps are published side by side, which makes right-sizing straightforward.
• Procurement options. Direct API or through Bedrock, Google Cloud or Foundry.
• Connects to other tools. Anthropic’s support docs say claude.ai plans can add custom connectors to remote MCP servers.

Limits and risks

• Text out only. The current models return text; they do not generate images. Pair Claude with an image tool when you need visuals.
• Prices and lineups change. The table below is a dated snapshot. Re-check it before you budget.
• Model vs harness. Practitioners such as Ethan Mollick argue that the tooling around a model matters as much as the model itself. Christopher Penn argues that for most text work, poor output is a prompt or setup problem, not a vendor problem. Treat these as informed views, not measurements.