Can I call Claude with the OpenAI SDK?

Yes. Point the OpenAI SDK at https://api.cognocient.com/v1 with your sk-cog- proxy key and set model to a Claude model name such as claude-sonnet-4-6. Cognocient AI translates the request to Anthropic's Messages API and the response back to OpenAI format, including streaming and tool calls.

How do I use the native Anthropic SDK through a proxy?

Set base_url to https://api.cognocient.com (no /v1; the SDK adds /v1/messages) and api_key to your sk-cog- proxy key. Requests are passed to Anthropic unchanged, so beta headers and new features work immediately.

Which route should I use for Claude?

Use the native Anthropic SDK route for Anthropic-specific features (prompt caching controls, extended thinking, PDFs, betas) and for Claude Code. Use the OpenAI SDK route when you want one client for every provider, or need guardrails, response caching, routing rules or provider failover, which run on that route only.

Is Claude prompt caching costed correctly?

Yes. Cache reads and cache writes (5-minute and 1-hour) are read from Anthropic's usage block and priced at their own rates, so a cached call shows its real, lower cost.

Does tool calling work with Claude through the OpenAI SDK?

Yes, streamed and non-streamed. OpenAI tools, tool_choice (auto, required, none or a named function) and parallel_tool_calls=false are translated to Anthropic's equivalents, and Claude's tool_use blocks come back as OpenAI tool_calls.

Claude API proxy: call Anthropic through the OpenAI SDK or the native Anthropic SDK

Route Claude calls through an LLM gateway two ways: the OpenAI SDK (change the model name to claude-*) or the native Anthropic SDK passthrough. Per-feature cost tracking, prompt-cache-aware costing and budgets enforced before the call.

Claude reaches Cognocient AI two ways. Both use the same proxy key, the same budgets and the same attribution headers, and both show up in the same dashboard.

	OpenAI SDK route	Native Anthropic SDK route
Base URL	`https://api.cognocient.com/v1`	`https://api.cognocient.com` (no `/v1`)
Request format	OpenAI Chat Completions, translated for you	Anthropic Messages API, passed through unchanged
Best for	One client for every provider; switching models by name	Anthropic-only features, betas, Claude Code
Budgets, Emergency Freeze, attribution, cost tracking	✅	✅
Prompt guardrails, response caching, routing rules, failover	✅	❌

Setup

Add your Anthropic API key in Settings → AI Providers, then pick a route.

The SDKs also read ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY from the environment, which is how Claude Code is pointed at the gateway.

The OpenAI SDK route: what is translated

Any model name starting with claude- is sent to Anthropic. The gateway converts the request to the Messages API and the reply back to the OpenAI shape:

OpenAI field	Becomes
`system` messages	The `system` prompt
`max_tokens` or `max_completion_tokens`	`max_tokens` (4,096 if neither is set)
`temperature`, `top_p`	Same
`stop` (string or list)	`stop_sequences`
`tools`	Anthropic tool definitions, on streamed and non-streamed calls
`tool_choice`: `auto` / `required` / `none` / `{"function": {"name"}}`	`auto` / `any` / `none` / `tool`
`parallel_tool_calls: false`	`disable_parallel_tool_use`
`role: "tool"` messages	`tool_result` blocks
`image_url` content parts	Image blocks (https URLs and base64 `data:` URLs)
Streaming	Anthropic SSE converted to OpenAI chunks, tool calls included, with a final usage chunk

response_format (JSON mode) is not translated. For structured output from Claude on this route, use a forced tool call (tool_choice with a function name), which is what LangChain's with_structured_output does.

The native route: passthrough

On https://api.cognocient.com/v1/messages your request body and anthropic-* headers reach Anthropic unchanged, so new models, beta headers, extended thinking, PDFs and prompt-caching controls work the day Anthropic ships them. Budgets, the Emergency Freeze, attribution headers and cost tracking apply. Prompt guardrails, response caching, routing rules and provider failover don't, because they need to read or rewrite the request in OpenAI form.

Prompt caching and cost

Anthropic reports cache reads and cache writes separately, and they are priced differently: reads well below the input rate, writes above it (with a higher rate for 1-hour writes). Cognocient AI prices each part at its own rate, so a call that hits the cache shows its real cost, and Context Tax (Growth plan) can tell you which features would benefit from caching their system prompt.

Claude on AWS Bedrock

Claude models hosted on Bedrock use Bedrock model IDs (anthropic.claude-..., or us. / eu. / global. inference profiles) and your AWS credentials. See AWS Bedrock.