Anthropic (direct)¶

For Claude-first users who want the latest model snapshots without waiting for them to cycle through Bedrock or a cloud provider. Calls go straight to api.anthropic.com.

Koji's Anthropic adapter targets the native Messages API (POST /v1/messages), not an OpenAI-compatible shim — the request shape and auth headers differ enough that routing Claude through the generic OpenAI adapter simply 404s.

1. Get an API key¶

Sign in at console.anthropic.com.
Settings → Billing — add credits. Anthropic uses a prepaid credit model rather than monthly billing by default.
API keys → Create Key. Copy the key — only shown once. Starts with sk-ant-.

2. Add the provider in Koji¶

Settings → Model Providers → Add provider.

Field	Value
Name	e.g. `anthropic-prod`
Provider	`Anthropic`
Base URL	`https://api.anthropic.com` (prefilled)
API key	`sk-ant-...` from step 1
Default model	See cheat sheet below

Save. Koji sends the key as x-api-key and pins the Messages API version to 2023-06-01.

3. Models¶

Anthropic uses date-suffixed model IDs. Pick a specific snapshot — Anthropic deprecates older snapshots on published timelines, so "latest" aliases are best avoided in config.

Model	Use it for
`claude-3-5-sonnet-20241022`	Default for extraction. Best accuracy/cost balance.
`claude-3-5-haiku-20241022`	High-throughput, simple-schema extraction. ~3× cheaper than Sonnet.
`claude-3-opus-20240229`	Most complex documents; slowest and most expensive.

Deprecation cadence: Anthropic typically announces ~6 months ahead of a snapshot cut-off. When a model you're using gets a deprecation notice, add a new provider with the newer snapshot, benchmark with koji bench, and flip the pipeline.

JSON mode — how it works¶

The Anthropic Messages API has no native equivalent to OpenAI's response_format: json_object. Koji's adapter emulates JSON mode by appending a short instruction to the user prompt:

Respond with ONLY a JSON object.

In practice this is reliable for schema extraction — Claude is good at obeying format instructions when the task is well-scoped. But there is one failure mode to know about:

max_tokens truncation

Anthropic requires max_tokens on every request. Koji defaults to 4096. For schemas that produce long nested arrays (e.g. invoices with 100+ line items, transcripts with many speakers), the model can truncate mid-JSON, and the json.loads(...) in the extract step then fails. If you see JSON-parse errors on Anthropic but not on OpenAI for the same schema + document, raise max_tokens — either per-pipeline or via a custom extraction step.

Rate limits¶

Anthropic scales limits by usage tier (Build, Scale, Enterprise). On lower tiers you can saturate with modest parallelism:

Tier 1: ~50 RPM, 40,000 ITPM (input tokens/min) per model
Tier 4: ~4,000 RPM, 2,000,000 ITPM

Check your organisation's rate limits page for current numbers. Same remediation as OpenAI — lower pipeline concurrency, wait for the tier to auto-promote, or contact Anthropic sales for a custom limit.

Known limitation: no tool use yet¶

Anthropic's tool use uses tool_use / tool_result content blocks rather than OpenAI's tool_calls array. Koji's Anthropic adapter does not yet translate between the two — schemas relying on tool-call-style field extraction will raise NotImplementedError on an Anthropic endpoint.

Workaround: for tool-use workloads that need Claude, route through Bedrock or an OpenAI-compatible proxy (LiteLLM, Portkey) that handles the translation. Native Anthropic tool-use is tracked on the internal roadmap.