Claude (Anthropic API)
Provider, cloud llm, by Anthropic
Assessment
The cloud engine for the work a home model cannot do: large refactors, hard reasoning, long documents. Treat each use as a deliberate choice, never a silent fallback, because the cost is money and the content leaves the house.
2026-10-09
Strengths
- The strongest reasoning and coding models available through an API, with large contexts
- Tool use, structured output, prompt caching and batch pricing
- The same models behind Claude Code, so a session can switch engines without changing tools
Limitations
- Usage-priced: a long agent session costs real money
- Everything sent leaves the network, which is a decision each time, not a default
Anthropic's API serves the Claude model family: as of 2026 the Claude 5 generation, with Opus for the hardest work, Sonnet for the everyday balance of quality and cost, and Haiku for speed.[1] The API supports tool use, JSON-shaped output, prompt caching that cuts the cost of repeated context, and batch processing at a discount.
What it is for
The jobs a local model handles badly: a refactor across a codebase, a long technical document, a design discussion that needs judgment, an agent run with many tool calls. Through a proxy that speaks the same interface as a local server, a coding agent can use Claude for one session and a home model for the next.
Things to know
- Cost is per token both ways. A long agentic session with large files in context adds up; prompt caching and keeping context tight matter.
- Egress. Code and documents sent to the API leave the home network. Where a rule says nothing leaves without a reason, the reason is per session, and a local model never escalates to this on its own.[2]
- Model ids move. Pin the id a project depends on and read the deprecation notices; the family naming (Opus, Sonnet, Haiku) is stable, the versions are not.
History
- 2026-10-09: article written and published.
- 2026-10-09: entry created.
Comments
Public comments on each entry are coming. Nothing is collected here yet.