Model Configuration

Baloo supports multiple LLM providers. AGENT_PROVIDER is global — every Baloo agent (primary review, FP verification, thread replies, fidelity, documentation drift, sync scope) uses that backend. Short names select a model tier on the provider; they are not a way to pick a different provider.

A provider/model string is accepted, but its provider must match AGENT_PROVIDER; a mismatch is rejected with a configuration error rather than silently overriding the provider for that agent. Bare model IDs (including Bedrock inference-profile IDs and ARNs) always run on the configured provider.

Model tiers

Short names	Tier	Max turns	Typical use
flash, haiku	Economy	10	FP verification, thread replies, simple PRs
sonnet, standard, gemini-pro	Standard	30	Default code reviews
opus, premium, gemini-3.1-pro	Premium	30	Complex / security-sensitive reviews

Resolved model IDs depend on AGENT_PROVIDER:

Provider	Economy	Standard	Premium
anthropic	claude-haiku-4-5-20251001	claude-sonnet-5	claude-opus-5-5
google	gemini-3.5-flash-lite	gemini-3.8-flash	gemini-3.1-pro-preview
amazon-bedrock	us.anthropic.claude-haiku-4-5-20251001-v1:0	us.anthropic.claude-sonnet-5	us.anthropic.claude-opus-5-5
openai	gpt-6-luna	gpt-6-sol	gpt-6-astra
databricks	system.ai.claude-haiku-4-5	system.ai.claude-sonnet-4-6	system.ai.claude-opus-4-6

Anthropic (and matching Bedrock Claude) tiers intentionally stay on Haiku 4.5 / Sonnet 4.6 / Opus 4.6 — a well-tested set. Newer Claude generations can be opted into later via bare model IDs or provider/model strings.

Choosing a Model

• Economy (flash, haiku) — Good for simple PRs (docs, deps, configs). Fast and cheap. Also used internally for FP verification and thread replies.
• Standard (sonnet, standard, gemini-pro) — The default. Handles most code reviews well. Best cost/quality balance.
• Premium (opus, premium, gemini-3.1-pro) — Best for complex PRs with deep logic, security-sensitive code, or architectural changes.

Switching Providers

Provider selection is all-or-nothing: AGENT_PROVIDER applies to every agent, and short names are tiers on it. Moving an existing deployment (for example Anthropic → Bedrock) is normally a one-variable change.

1. Make sure your model settings are tier short names. Anything set to a bare provider-specific ID (claude-sonnet-5) or an explicit provider/model string is passed through as-is and will not translate. Short names (sonnet, haiku, opus) travel across providers; the defaults already use them.

2. Change the provider. Either set the environment variable and restart:

AGENT_PROVIDER=amazon-bedrock

…or, when DATABASE_ENABLED=true, change it on the dashboard Settings page with no restart. See Runtime Overrides.

3. Add that provider's credentials — see API Keys.

4. Verify. The Models in use table on the Settings page should show the new provider for every role, and Test connection should pass.

Amazon Bedrock

For a full walkthrough (auth methods, sandbox caveats, verification, troubleshooting) see the Amazon Bedrock Setup guide. In short:

pi's provider token is amazon-bedrock. Point Baloo at it with:

AGENT_PROVIDER=amazon-bedrock
AGENT_MODEL=sonnet
# or a specific inference profile / ARN:
# AGENT_MODEL=us.anthropic.claude-sonnet-5
# AGENT_MODEL=amazon-bedrock/us.anthropic.claude-sonnet-5
AWS_REGION=us-east-1

With AGENT_PROVIDER=amazon-bedrock, short names such as haiku (FP/thread defaults) and sonnet (primary default) resolve to the Bedrock tier IDs above. Override with a bare Bedrock model ID or application inference profile ARN when your account uses different regional prefixes (global., eu., …).

Auth (pick one): IAM access keys (+ optional session token), AWS_BEARER_TOKEN_BEDROCK, AWS_PROFILE, IRSA (AWS_WEB_IDENTITY_TOKEN_FILE + AWS_ROLE_ARN), or ECS/EC2 instance roles. Baloo allowlists these AWS env vars into the sandboxed pi subprocess and bind-mounts IRSA/credential files when their paths are set. See Configuration.

Cost estimation for Bedrock models uses the cost pi reports (Baloo's built-in pricing table is Anthropic-first-party only).

Databricks

For a full walkthrough (token scopes, model availability, sandbox caveats, troubleshooting) see the Databricks Setup guide. In short:

pi has no native Databricks provider, so Baloo generates a models.json registering one against the workspace AI Gateway and points pi at it with PI_CODING_AGENT_DIR. Point Baloo at it with:

AGENT_PROVIDER=databricks
AGENT_MODEL=sonnet
# or a specific Unity Catalog model service:
# AGENT_MODEL=system.ai.claude-sonnet-4-6
DATABRICKS_HOST=https://dbc-xxxxxxxx-xxxx.cloud.databricks.com
DATABRICKS_TOKEN=dapi...

Model IDs are Unity Catalog model services (system.ai.claude-*); the older flat databricks-claude-* names return 501 NOT_IMPLEMENTED. Availability is per workspace — an unprovisioned model reports a rate limit of 0.

Cost is reported as $0 for Databricks: it bills DBUs at a per-contract rate, so there is no correct USD-per-token constant to hardcode. See Cost reporting.

Configuration

# Provider for all agents
AGENT_PROVIDER=anthropic

# Primary review tier (short name on that provider)
AGENT_MODEL=sonnet

# Or a full provider/model string (provider must match AGENT_PROVIDER)
AGENT_MODEL=anthropic/claude-sonnet-5

# Premium model for highest quality
AGENT_MODEL=opus

When DATABASE_ENABLED=true, AGENT_PROVIDER, AGENT_MODEL, and PI_THINKING_LEVEL can also be changed at runtime from the dashboard Settings page without restarting. See Runtime Overrides.

API Keys

Each provider needs its own credentials:

Provider	Environment Variable / Auth
Anthropic	ANTHROPIC_API_KEY
Google	GEMINI_API_KEY
OpenAI	OPENAI_API_KEY
Amazon Bedrock	AWS credentials / IRSA / bearer token (see Amazon Bedrock)
Databricks	DATABRICKS_HOST + DATABRICKS_TOKEN (see Databricks Setup)

Thinking Level

Controls the depth of reasoning the model uses:

PI_THINKING_LEVEL=medium  # off, minimal, low, medium, high, xhigh, max

Higher thinking = better analysis but slower and more expensive. medium is the default and recommended for most use cases.

Cost Estimates

Approximate cost per review (typical 5-file PR):

Model	Cost per Review
flash	~$0.005
haiku	~$0.01
sonnet	~$0.03–0.08
opus	~$0.15–0.40