Anthropic

Anthropic is the home of the Claude family, and claude-fable-5 — the first of the Claude 5 line — is the strongest coding and agentic model in Stella's catalog. If you want the best worker money currently buys, or a top-shelf judge over a cheaper worker, this is the provider. Stella speaks the Anthropic Messages dialect natively against https://api.anthropic.com.

Setup

export ANTHROPIC_API_KEY="sk-ant-..."

That is the whole setup — Anthropic sits second in auto-detection, so with the key exported Stella picks it up on its own (only a ZAI_API_KEY outranks it). Prefer a file? Store the key in ~/.config/stella/credentials.toml, or rename the env var Stella reads with providers.anthropic.api_key_env — see overriding the key source.

Selecting models

The provider id is anthropic; the default model is claude-fable-5.

stella --model anthropic/claude-fable-5 run "refactor the session store"
export STELLA_MODEL=anthropic/claude-fable-5   # pin for the whole shell

Custom base URL

Anthropic-compatible gateways and proxies — a LiteLLM deployment, a corporate egress proxy — work by overriding the base URL; the wire dialect stays Anthropic Messages:

{
  "providers": {
    "anthropic": {
      "base_url": "https://llm-gateway.internal.example.com"
    }
  }
}

For a one-off run, --base-url (env form STELLA_BASE_URL) does the same and outranks the settings value. User scope shown deliberately: in a project's .stella/settings.json, base_url is ignored unless STELLA_TRUST_PROJECT=1 — see the project trust boundary.

Reasoning and sampling

Stella's reasoning setting maps to Anthropic extended thinking, with a thinking-token budget tiered by effort — a higher effort buys a larger budget. One API rule to know: while thinking is enabled, temperature is omitted from the request — the Messages API does not accept both, and Stella complies rather than erroring. Set reasoning: "off" for an agent if you need temperature control back.

Prompt caching

Prompt caching is first-class on this provider: Stella meters both cache reads and cache writes, so the telemetry you see in the Observatory reflects the true bill. Cached input is billed at $0.30/Mtok against a $3.00 list input price — a 10x discount on the byte-stable prefix the context engine maintains — which is why long claude-fable-5 runs cost far less than the headline price implies.

In the pipeline

claude-fable-5 is the reference worker for hard multi-file changes and a superb judge — the max-quality profile uses it as the worker under a gpt-5.5 judge, while the balanced profile flips it into the judge seat over a glm-5.2 worker. Route it per agent with agent_engine_config.

Catalog pricing

List prices from Stella's seed catalog — verify with Anthropic; your local catalog re-syncs from models.dev via stella models refresh. See the Model guide for the full cross-provider table.