Anthropic

Run Claude models through the Anthropic Messages API. The same provider also reaches Claude on Google Cloud Vertex AI and on Amazon Bedrock.

Setup

pip install "10xgraph[anthropic]"

Get an API key from console.anthropic.com and export it:

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

Or add it to a .env file:

ANTHROPIC_API_KEY=sk-ant-...

An unset ANTHROPIC_API_KEY is not fatal. The SDK also resolves ANTHROPIC_AUTH_TOKEN, an ant auth login profile, and workload identity federation, so 10xGraph logs an informational message and lets the SDK try.

Basic usage

from tenxgraph.core.graph import Agent

agent = Agent(
    model="claude-opus-5",
    provider="anthropic",
    system_prompt=[{"role": "system", "content": "You are a helpful assistant."}],
)

provider is optional here: model names starting with claude- or anthropic. resolve to the Anthropic provider, and so does an explicit anthropic/ or claude/ prefix on the model string.

Backends

Google switches to Vertex AI with a boolean use_vertex_ai. Anthropic reaches three distinct backends, so its selector is the string anthropic_backend, passed through llm_kwargs.

# Direct Claude API
agent = Agent(model="claude-opus-5")

# Google Cloud Vertex AI — bare model id
agent = Agent(model="claude-opus-5", anthropic_backend="vertex")

# Amazon Bedrock — model ids keep their `anthropic.` prefix
agent = Agent(model="anthropic.claude-opus-5", anthropic_backend="bedrock")

Region, project, and AWS credentials are resolved by the Anthropic SDK from the environment unless you pass them explicitly through llm_kwargs. Passing use_vertex_ai=True without an explicit anthropic_backend also selects the Vertex backend, so the flag means the same thing for Claude as it does for Gemini.

The Bedrock client is AsyncAnthropicBedrockMantle, the Messages-API endpoint. The plain AsyncAnthropicBedrock client is the legacy InvokeModel path and is deliberately not used.

Output types

output_type accepts "text" and "json". Anything else raises a ValueError at construction time. Claude’s Messages API generates text and tool calls only; there is no image, audio, or video generation endpoint, so those output types belong to other providers.

Reasoning

reasoning_config={"effort": ...} maps onto thinking={"type": "adaptive"} plus output_config={"effort": ...}.

agent = Agent(
    model="claude-opus-5",
    reasoning_config={"effort": "high"},
)

budget_tokens and thinking_budget are not sent. Current Claude models return a 400 for an explicit thinking budget alongside the model’s own adaptive control; 10xGraph logs a warning and drops them. Use effort to control depth.

max_tokens

Anthropic requires max_tokens on every request, so 10xGraph supplies a default when you do not:

Streaming gets the larger default because a high max_tokens on a non-streaming request risks an HTTP timeout. Pass max_tokens explicitly to override either one.

Sampling parameters

Several current models reject temperature, top_p, and top_k with a 400. 10xGraph strips those three keys before the request for claude-fable-5, claude-mythos-5, claude-opus-5, claude-opus-4-8, claude-opus-4-7, and claude-sonnet-5 (Bedrock’s anthropic. prefix is stripped before the check). Older Claude models still accept them, so this is a per-model set rather than a blanket strip.

Prompt caching

Set anthropic_cache to place cache_control breakpoints on the stable prefix of the request.

agent = Agent(
    model="claude-opus-5",
    anthropic_cache=True,                                # ephemeral, default TTL
    # anthropic_cache={"type": "ephemeral", "ttl": "1h"} # extended retention
)

Caching is a prefix match. The render order is tools → system → messages, so the breakpoint goes at the end of the stable prefix — the last tool definition when tools are present, and the last system block — leaving volatile per-request messages after it. Any byte change invalidates everything after the change.

The minimum cacheable prefix is roughly 1024 tokens; a shorter prefix silently does not cache. Confirm with usage.cache_read_input_tokens on the response.

Batch requests

AnthropicBatch wraps the Message Batches API for offline, high-volume work.

from tenxgraph.core.llm import AnthropicBatch

batch = AnthropicBatch(model="claude-haiku-4-5")
batch.add("row-1", [{"role": "user", "content": "Summarise: ..."}])
batch.add("row-2", [{"role": "user", "content": "Summarise: ..."}])

batch_id = await batch.submit()
results = await batch.wait(batch_id)      # keyed by custom_id
print(results["row-1"].text)

Results arrive in any order, so they are keyed by custom_id throughout — indexing by position is the classic way to silently mismatch a batch. status(batch_id) polls once without blocking, and results(batch_id) collects a batch you already know has ended. Messages use 10xGraph’s internal dialect and go through the same translation as a live call. OpenAIBatch exposes the same surface for the OpenAI provider.

Environment Variables

One credential source must resolve. 10xGraph only reads ANTHROPIC_API_KEY itself; everything else is left to the SDK’s own resolution. Vertex and Bedrock backends use their platform’s standard credential chain instead.

Common Errors

Related docs

• Providers
• Installation extras matrix