OpenAI API vs Anthropic API: Comparison and llm-connector Compatibility Audit

1. API Comparison

1.1 Interface Design

Dimension	OpenAI Chat Completions	Anthropic Messages
Message model	messages array with role/content per item	messages array + separate system field
Roles	user, assistant, system, tool	user, assistant (system in separate field)
Content format	content as string or content blocks array	Always content blocks array
Endpoint	/v1/chat/completions	/v1/messages
Style	Mature, many parameters, broad compatibility	Clean structure, XML support, consistent output

1.2 Authentication

Dimension	OpenAI	Anthropic
Primary auth	Authorization: Bearer {api_key}	x-api-key: {api_key}
Version header	None required	anthropic-version: 2023-06-01
Azure variant	api-key + api-version	Bearer via Vertex/Bedrock proxy

1.3 Request/Response Format

Field	OpenAI	Anthropic
Model	model	model
Max tokens	max_tokens optional	max_tokens required
Temperature	temperature	temperature
Output	choices[] with message	Single content[], no choices
Token stats	usage.prompt_tokens / completion_tokens	usage.input_tokens / output_tokens
Stop reason	finish_reason	stop_reason

1.4 Error Codes

HTTP	OpenAI	Anthropic
400	invalid_request_error	invalid_request_error
401	authentication_error	authentication_error
403	permission_error	permission_error
429	rate_limit_error	rate_limit_error
500+	Server error	Server error
Special	context_length_exceeded (code)	Detected via message content

1.5 Streaming

Dimension	OpenAI	Anthropic
Protocol	SSE	SSE
Events	data: lines, choices[0].delta	message_start, content_block_delta, message_delta, message_stop
Structure	Incremental delta	Block-level delta + metadata
tool_calls	Incremental accumulation by index	Parse content_block_delta for tool_use

1.6 Function Calling

Dimension	OpenAI	Anthropic
Definition	tools[].function (name/description/parameters)	tools[].input_schema (JSON Schema)
Calls	tool_calls[] with id/function/name/arguments	tool_use blocks in content
Results	tool role + tool_call_id	tool_result in user role
tool_choice	auto/none/required/function	No direct equivalent, prompt-based control

1.7 Multimodal

Dimension	OpenAI	Anthropic
Image input	image_url (URL or base64)	image (base64, type: image/source)
Blocks	text, image_url	text, image
Formats	URL and base64	Primarily base64

1.8 Rate Limits

Both use RPM/TPM, per-model; OpenAI exposes x-ratelimit-* headers; both return 429 with Retry-After behavior.

1.9 Pricing

Both charge by token; input/output rates differ. llm-connector does not implement billing.

1.10 SDK Ecosystem

Both provide official Python and TypeScript SDKs. llm-connector offers a unified Rust layer.

---

2. Version Evolution (since 2022)

OpenAI

Date	Milestone
2022	Chat Completions release
2023	Function calling, Vision, GPT-4 Turbo
2024	o1 reasoning, structured outputs, reasoning_content
2025-03	Responses API released, recommended for new projects
2025-05	MCP, Code Interpreter, Image generation, Background mode
2026-08	Assistants API planned sunset

Anthropic

Date	Milestone
2023	Messages API, anthropic-version: 2023-06-01
2024	Tool use (function calling), extended thinking
2025	Version updates, built-in tools
2026-02	Built-in tools GA

---

3. llm-connector Compatibility Audit

3.1 OpenAI Coverage

Endpoint/Feature	Supported	Notes
POST /chat/completions	✓
└ model, messages, temperature, max_tokens, top_p	✓
└ frequency_penalty, presence_penalty	✓
└ stream, tools, tool_choice, response_format	✓
└ stop, logit_bias, user, seed	✗	Not mapped to OpenAIRequest
GET /models	✓
Streaming	✓	delta.content, tool_calls accumulation, reasoning_content

Unmapped (ChatRequest → OpenAIRequest): stop, logit_bias, user, seed

3.2 Anthropic Coverage

Endpoint/Feature	Supported	Notes
POST /v1/messages	✓
└ model, max_tokens, messages, system, temperature, top_p, stream	✓
└ tools, tool_choice, response_format	✗	Not implemented
Streaming	✓	message_start, content_block_delta, message_delta
tool_use streaming	✗	Tool results converted to user text, no real tool support
models list	✗	models_endpoint returns None

3.3 Gap Matrix and Priority

ID	Gap	Type	Priority	Description
D1	Anthropic tools	Missing	P0	No function calling
D2	Anthropic tool streaming	Missing	P0	tool_use not parsed
D3	OpenAI stop/logit_bias/user/seed	Unmapped	P1	Affects advanced usage
D4	Anthropic models endpoint	Missing	P1	fetch_models unsupported
D5	OpenAI logprobs	Unmapped	P2	Low priority
D6	response_format (Anthropic)	Missing	P2	Limited structured output
D7	Tool role → user text	Semantic mismatch	P1	Tool results treated as plain user messages

---

4. Adaptation Plan

4.1 New Modules

• OpenAI Responses Protocol: Add OpenAIResponsesProtocol for /v1/responses; see OpenAI Responses API Analysis
• Optional: Anthropic tools adapter for input_schema ↔ function.parameters mapping

4.2 Refactoring

Module	Changes
anthropic.rs	Add tools, tool_choice in build_request; tool_use in parse_response / parse_stream_response
openai.rs	Optionally add stop, logit_bias, user, seed to OpenAIRequest
types/request.rs	Add or extend ResponsesRequest if adopting Response API

4.3 Integration Tests

• openai_tools_e2e: Non-streaming + streaming tool calls
• anthropic_tools_e2e: Anthropic tool use (requires API key or mock)
• anthropic_stream_tool_use: Anthropic streaming tool_use parsing
• openai_stop_sequences: stop parameter behavior
• responses_api_migrate: Response API basic request/response (once implemented)

4.4 Documentation and Examples

• README: Anthropic tools support
• Providers: Anthropic capability matrix
• Tools: Anthropic tool use example
• examples/: Add anthropic_tool_use.rs if implemented

4.5 Milestones and Acceptance

Milestone	Scope	Acceptance
M1	Anthropic tools (P0)	≥95% unit, ≥90% integration, zero P0 defects
M2	OpenAI unmapped fields (P1)	Same
M3	OpenAI Responses API	Same + dedicated Response API tests
M4	Docs and examples	All above updates done