Anthropic

When using Anthropic models via AWS Bedrock, llm-exe will make POST requests to https://bedrock-runtime.us-west-2.amazonaws.com/model/{MODEL_ID}/invoke.

Setup

Anthropic Chat

const llm = useLlm("amazon:anthropic.chat.v1", {
  model: "anthropic.claude-sonnet-4-v2:0",  // This is the model id from Bedrock
});

Bedrock Anthropic Options

In addition to the generic options, the following options are available for Anthropic models on Bedrock.

Option	Type	Default	Description
model	string	—	The Bedrock model id, cross-region inference profile id, or a full ARN. Must be specified. See the model identifiers note below, and the AWS Bedrock Docs
maxTokens	number	10000	Maps to max_tokens. Raised to 65536 when unset on the escalated effort path (Opus 4.7 / 4.8 high). See the effort note below.
topP	number	undefined	Maps to top_p. Dropped for reject models, or < 0.95 under effort.
effort	string	undefined	Maps to output_config.effort + thinking. See Anthropic provider.
awsRegion	string	undefined	AWS Region. Can be set via AWS_REGION environment variable
awsSecretKey	string	undefined	AWS Secret Key. Can be set via AWS_SECRET_ACCESS_KEY environment variable
awsAccessKey	string	undefined	AWS Access Key. Can be set via AWS_ACCESS_KEY_ID environment variable

NOTE

The Bedrock Anthropic provider maps a subset of the direct Anthropic provider options. temperature, topK, stopSequences, metadata, and serviceTier are not mapped for the Bedrock variant at this time.

effort and sampling params: effort maps to output_config.effort and adaptive/extended thinking, identical to the direct provider, including the Opus 4.7 / 4.8 high -> xhigh escalation, which raises the default max_tokens to 65536 when you do not set it. topP is dropped for the models that 400 on sampling parameters (Opus 4.7, 4.8, and 5; Sonnet 5; Fable 5).

Timeout on the escalated path: the raised 65536 ceiling can let an xhigh run exceed the default 30000ms timeout (llm-exe does not stream), and timeouts are retried (default numOfAttempts 2), costing a second billed attempt. Raise timeout when using effort: "high" on Opus 4.7 / 4.8.

Because effort enables thinking, and Anthropic disallows sampling parameters while thinking is active, llm-exe drops topP below 0.95 whenever effort enables thinking (and drops temperature / topK on the direct provider, which maps them). Pass topP >= 0.95 if you need it. This applies to the direct provider as well.

Model identifiers: model accepts three forms, and llm-exe URL-encodes it into the invoke path, so identifiers containing / or : route correctly:

// 1. Foundation model id
useLlm("amazon:anthropic.chat.v1", { model: "anthropic.claude-sonnet-4-v2:0" });

// 2. Cross-region inference profile id — the usual form for newer models
useLlm("amazon:anthropic.chat.v1", { model: "us.anthropic.claude-sonnet-4-v2:0" });

// 3. Full ARN
useLlm("amazon:anthropic.chat.v1", {
  model: "arn:aws:bedrock:us-east-1:123456789012:inference-profile/us.anthropic.claude-sonnet-4-v2:0",
});

The geographic prefix (us., eu., apac., us-gov.) is understood, so a cross-region profile is gated for effort exactly like the bare foundation model id.

ARN types and effort (easy to miss): the model-specific behavior below — effort / thinking, the Opus 4.7 / 4.8 escalation, and the topP drop — is selected by matching the Claude model name. llm-exe can recover that name from an inference-profile/ or foundation-model/ ARN, because those embed the model id. The other ARN types (provisioned-model/, custom-model/, application-inference-profile/) carry an opaque id, so no gate matches: the request is still dispatched, but effort is silently ignored and topP is not dropped. If you point at a provisioned throughput or application inference profile ARN and need this behavior, pass the tuning options explicitly rather than relying on effort.

Forced tool_choice: functionCall: "any" (or a named tool) is incompatible with extended thinking (the 4.5 enabled path) on both providers, and llm-exe throws a clear error for it. Note Bedrock additionally rejects a forced tool_choice with adaptive thinking — which the direct API accepts — so combining effort with functionCall: "any" on an adaptive model returns a 400 from Bedrock; use functionCall: "auto" there.