Feature Matrix

57 capabilities
across every official SDK

What each official SDK actually exposes -- not what the marketing pages claim. Every cell is checked against the pinned source of that provider's own SDK, and clicking one shows you the exact lines it was verified against.

281

cited sources

213

inline snippets

API ≠ SDK gaps

unverified cells

How this was built

Nothing is asserted from memory. Every “supported” claim cites a file and line in that provider's own SDK, at a pinned version. A build check fails if a citation stops resolving.
Accepted is not honoured. Where a parameter is only meaningful if it changes behaviour, we tested it rather than trusting a 200. top_k is accepted by three providers and actually applied by one.
Two different questions. A capability can exist on a provider's API while its official SDK doesn't expose it — xAI ships a voice API its Python SDK has no module for. Those cells are marked API.
Gaps are shown, not hidden. Rows where we're weaker than an official SDK are in here too — filter by Our gaps.

Showing 57 of 57 features · click any cell for the evidence

●Supported◐Partial◑Beta / gated○Not supported–Not applicableAPIprovider has it, SDK doesn't expose it

Feature	ORXA llm-sdk	Anthropic	OpenAI	Google	xAI
Core completion The shape of a single request/response: what you send, how the turn is framed, and what comes back.
Single-prompt completion core.prompt · example
System prompt core.system · example
Multi-turn conversation core.multi-turn · example
Output length cap core.max-tokens
Sampling parameters The knobs that shape token selection. The interesting part is not who has them, but who SILENTLY IGNORES them.
Temperature / top-p sampling.temperature
Top-k sampling.top-k
Presence / frequency penalties sampling.penalties
Stop sequences sampling.stop
Deterministic seed sampling.seed
Streaming Incremental delivery of a turn. Every provider streams text; they diverge sharply on what ELSE is streamable and how events are shaped.
Streaming text streaming.text · example
Rich stream events streaming.events
Tools / function calling Letting the model call your code. Universally supported, and universally spelled differently.
Function tool definition tools.define · example
Tool choice control tools.choice · example
Parallel tool calls tools.parallel · example
Automatic tool-execution loop tools.loop · example
Structured output Getting a typed object instead of prose. The gap here is less about support and more about what happens when the model gets it wrong.
JSON-Schema constrained output structured.schema · example
Typed parse + failure handling structured.parse · example
Vision & documents Non-text INPUT. Everyone takes images; they diverge on PDFs, on whether a URL can be passed directly, and on how big an input may be.
Image input vision.image · example
PDF / document input vision.document · example
Audio Speech in and speech out. The widest capability spread of any category — one provider has none of it.
Speech-to-text audio.stt · example
Text-to-speech audio.tts · example
Image generation Producing images rather than reading them. Note how differently the three providers model it: a dedicated endpoint, a normal completion, or both.
Generate an image imagegen.generate · example
Edit an existing image imagegen.edit
Files API Uploading content once and referencing it by id. Same idea everywhere, but the lifetime and the beta status differ sharply.
Upload & reference by id files.upload · example
Batch / async jobs Submit many requests for offline processing at a discount. Universally ~50% cheaper, and universally shaped differently.
Batch submission & polling batch.submit · example
Prompt caching Reusing computation across requests. The single most divergent category in the matrix: explicit markers, an opaque key, or a first-class resource.
Prompt / context caching caching.prompt · example
Server-side conversation state Letting the provider hold the transcript so each turn sends only what is new.
Server-held conversation state state.previous-response · example
Hosted (provider-run) tools Tools the PROVIDER executes server-side — you never run any code. The least converged area in the matrix: OpenAI, Google and xAI all offer hosted retrieval over your own corpus (vector stores / File Search Stores / Collections); Anthropic offers none. Tool NAMES and shapes differ everywhere.
Web search hosted-tools.web-search · example
Code execution hosted-tools.code-execution · example
Hosted retrieval / file search hosted-tools.file-search
MCP (Model Context Protocol) Connecting external tool servers. Two postures: the provider connects (hosted), or your client connects (local). Anthropic and Google both integrate the external `mcp` package for the local case; OpenAI ships hosted MCP only.
MCP client (you connect) mcp.client · example
Hosted MCP (the provider connects) mcp.hosted
Realtime / live Bidirectional low-latency sessions over a socket, usually voice. Only two providers have it, and they share almost no vocabulary.
Realtime session realtime.session · example
Embeddings Vector representations. Notable mainly for who does NOT have them.
Text embeddings embeddings.embed · example
Agent loop & orchestration Multi-step autonomous runs. IMPORTANT: for OpenAI and Google these live in a SEPARATE package (@openai/agents, google-adk) — not the SDK you install to call the model. Cells cite whichever package actually provides it.
Multi-step agent loop agent.loop · example
Guardrails agent.guardrails
Human-in-the-loop approval agent.approval
Context management Keeping a long conversation inside the window: measuring it, and compacting it when it grows.
Automatic context compaction context.compaction
Usage, tokens & cost Knowing what a call consumed and what it cost. Every provider reports tokens; TWO turn them into money by opposite routes - xAI’s API returns the ACTUAL billed cost, while we price a bundled catalog locally (and can therefore also estimate BEFORE sending).
Token usage reporting usage.tokens
Cost calculation usage.cost · example
Observability Seeing what happened: traces, spans, metrics and events.
Tracing / spans observability.tracing
Networking & reliability What happens between your process and the API: retries, timeouts, and whether anything protects you from your own concurrency.
Retries & timeouts networking.retry
Client-side rate limiting & queueing networking.queue
Errors & typing How failures are surfaced, and whether they mean the same thing across providers.
Typed error hierarchy errors.typed
Auth & configuration Keys, endpoints and per-request overrides.
API key & endpoint configuration auth.keys
Video generation Long-running video jobs. All three that offer it are async (submit → poll → download), and all three shape it differently.
Generate video videogen.generate
Moderation & safety Screening content. One provider has a dedicated classifier, one has request-level safety controls, one has neither.
Content moderation moderation.screen
Model catalog & discovery Finding out which models exist, what they cost, and what they can do. Every SDK lists ids; none of them tells you anything else.
List available models catalog.list · example
Pricing & capability metadata catalog.capabilities
Routing & fallback Sending a call somewhere else when the first choice fails or costs too much.
Fallback to another model routing.fallback · example
Runtime & environment Where the library can actually run, and what it drags in.
Browser support runtime.browser
Runtime dependencies runtime.dependencies
Reasoning / thinking Controls over a model's internal reasoning: whether it runs, how hard, how much comes back, and whether it persists across turns.
Enable reasoning reasoning.enable · example
Reasoning effort reasoning.effort · example
Reasoning visibility reasoning.visibility · example
Cross-turn reasoning persistence reasoning.context
Reasoning execution mode reasoning.mode