# Alternatives to OpenAI Realtime API > Source: https://theagentsindex.com/openai-realtime-api/alternatives (current datasheets, evidence-ranked) Candidates share the baseline category or are previously researched peers. Requirements, preserved strengths, explicit trade-offs and research coverage decide the order; paid tier does not. ## Shortlist ### 1. ElevenLabs Agents Hosted platform for building voice and chat agents, with telephony, SDKs and a choice of LLM. **Research coverage:** 35/35 criteria **Where it improves on OpenAI Realtime API:** - [Runs autonomously: yes — Agents run whole calls and chats unattended, including scheduled outbound campaigns, with explicit system tools to end a call or hand off to a human number.](https://elevenlabs.io/docs/eleven-agents/phone-numbers/batch-calls) - [Claude: yes — Anthropic models are natively selectable in agent settings; a few older Claude models drop out in EU data-residency workspaces.](https://elevenlabs.io/docs/eleven-agents/customization/llm) - [Open models: yes — Open-weight models run two ways: ElevenLabs-hosted Qwen in the native model list, or Llama and Mixtral through a custom LLM endpoint such as Together AI, Groq or SambaNova.](https://elevenlabs.io/docs/eleven-agents/customization/llm/custom-llm/together-ai) **Trade-offs against OpenAI Realtime API:** - [Admin controls: partial — Guardrails constrain what an agent may say and per-tool approval constrains what it may do, but Guardrails is in Alpha and enabling MCP servers is not restricted to admins.](https://elevenlabs.io/docs/eleven-agents/best-practices/guardrails) - [Audit log: partial — OCSF-schema logs over 100 administrative endpoints, pullable into a SIEM, but gated on Enterprise tier plus an API key holding audit_log_read.](https://elevenlabs.io/docs/overview/administration/workspaces/audit-logs) [View ElevenLabs Agents](https://theagentsindex.com/elevenlabs-conversational-ai) ### 2. AudioCodes Live Hub Self-service cloud platform for building voice AI agents and connecting them to real telephony channels. **Research coverage:** 33/35 criteria **Where it improves on OpenAI Realtime API:** - [Runs autonomously: yes — Voice AI Agents are designed to handle an entire inbound call end to end, answering questions and completing tasks with no human agent involved.](https://www.audiocodes.com/livehub) - [Claude: partial — Claude models are supported only as bring-your-own custom models configured with your own Anthropic API key; the pre-deployed model list carries OpenAI and Google models only.](https://techdocs.audiocodes.com/livehub/Content/AI-Agents/Bring%20your%20own%20models.htm) - [Open models: partial — Open-weight models such as gpt-oss and Gemma are reachable only through the bring-your-own-model path with your own Groq, Cerebras or OpenAI keys, not from the pre-deployed set.](https://techdocs.audiocodes.com/livehub/Content/AI-Agents/Bring%20your%20own%20models.htm) **Trade-offs against OpenAI Realtime API:** - [Agent permissions: partial — An agent calls any tool you configured without a per-call approval step, so control is exercised up front by choosing its tools; the MCP management connection does offer an explicit read-only mode.](https://techdocs.audiocodes.com/livehub/Content/AI-Agents/Tools.htm) [View AudioCodes Live Hub](https://theagentsindex.com/audiocodes-live-hub) ### 3. Synthflow Enterprise voice AI platform with in-house telephony for inbound and outbound call automation. **Research coverage:** 35/35 criteria **Where it improves on OpenAI Realtime API:** - [Runs autonomously: yes — Agents answer and place calls unattended, escalating to a human only on triggers you configure.](https://synthflow.ai/) - [Model choice: partial — You pick the model per agent, but only from Synthflow's list of GPT and Synthflow LLM options.](https://docs.synthflow.ai/general-configuration) - [In your editor: partial — No Synthflow editor extension; Claude, Claude Code and Cowork can drive the workspace over MCP from where a developer already works.](https://docs.synthflow.ai/anthropic) **Trade-offs against OpenAI Realtime API:** - [On the command line: no — The deployment channels are enumerated as telephony, WhatsApp, WebSocket, chat, API, Workflows, Zapier and GoHighLevel; Synthflow ships no command-line tool.](https://docs.synthflow.ai/deploy-your-agent) - [Open source: no — Subscriber Terms grant a non-transferable licence to use the software and forbid decompiling it.](https://docs.synthflow.ai/terms-conditions) - [Audit log: partial — API logs give an audit trail of every request an agent makes, with retention set by plan; no admin-action audit log is documented.](https://docs.synthflow.ai/logs) [View Synthflow](https://theagentsindex.com/synthflow-ai) ### 4. Vapi Developer platform for building, testing and deploying voice AI agents that make and receive phone calls. **Research coverage:** 35/35 criteria **Where it improves on OpenAI Realtime API:** - [Runs autonomously: yes — Agents take whole calls end to end and escalate to a human only when the configured path says so.](https://docs.vapi.ai/quickstart/introduction) - [Claude: yes — Claude is selectable as the assistant's LLM, on Vapi's integration or your own Anthropic account.](https://docs.vapi.ai/providers/model/anthropic) - [Open models: yes — Open-weight models reach the pipeline through Groq, Together AI, Mistral, DeepSeek and OpenRouter, or a self-hosted server.](https://vapi.ai/platform) **Trade-offs against OpenAI Realtime API:** - [Agent permissions: partial — Tools fire mid-call without human approval; a per-tool rejection plan is the mechanism that withholds an action.](https://docs.vapi.ai/tools/tool-rejection-plan) - [Audit log: partial — The dashboard keeps org-wide API, call and webhook logs and assistant version history; no member-action audit trail is documented.](https://docs.vapi.ai/debugging) - [Opt out of training: partial — The privacy policy reserves call content for model training and names no toggle; the $1,000/month ZDR add-on is the only documented way to stop it being kept at all.](https://vapi.ai/privacy) [View Vapi](https://theagentsindex.com/vapi) ### 5. Deepgram Voice Agent API One WebSocket API that runs speech-to-text, LLM orchestration and text-to-speech for real-time voice agents. **Research coverage:** 35/35 criteria **Where it improves on OpenAI Realtime API:** - [Runs autonomously: yes — The API runs the whole listen-think-speak loop live, handling interruptions and calling functions to take action without a human in the conversation.](https://deepgram.com/product/voice-agent-api) - [Claude: yes — Anthropic is a Deepgram-managed provider, so no endpoint of your own is needed; Claude Sonnet models bill at the Advanced tier and Claude Haiku at Standard.](https://developers.deepgram.com/docs/voice-agent-llm-models) - [Open models: yes — Open-weight LLMs are listed as supported models: gpt-oss-20b through Groq, which requires you to supply the endpoint, and an NVIDIA Nemotron model that Deepgram manages. Deepgram's own speech and voice weights are not distributed.](https://developers.deepgram.com/docs/voice-agent-llm-models) **Trade-offs against OpenAI Realtime API:** - [Multi-agent: partial — Handoff between specialised agents is a documented pattern with a reference implementation you run yourself: your orchestrator opens a fresh Voice Agent session per agent and summarises context between them. The API itself has no built-in agent-to-agent coordination.](https://developers.deepgram.com/docs/multi-agent-architecture) - [Agent permissions: partial — What the agent may do is bounded by the function definitions you send in Settings, and you choose whether each function executes client-side or against an endpoint you host. No approval or confirmation gate is documented: once a function is defined, the model may call it mid-conversation.](https://developers.deepgram.com/docs/voice-agents-function-calling) - [Admin controls: partial — Role-based control exists over the account, not over the agent: owner, admin and member roles plus scoped API keys govern who may read usage, create keys or change billing. What the agent itself may say or call is bounded by your prompt and function definitions, with no admin-side policy engine documented.](https://developers.deepgram.com/guides/deep-dives/managing-projects) [View Deepgram Voice Agent API](https://theagentsindex.com/deepgram-voice-agent-api) ### 6. Cartesia Real-time speech models and a managed runtime for production phone and web voice agents. **Research coverage:** 35/35 criteria **Where it improves on OpenAI Realtime API:** - [Runs autonomously: yes — A deployed agent runs a whole call unattended, handling turn-taking, interruptions and tool calls; it transfers to a human only when a configured condition matches.](https://docs.cartesia.ai/agents/introduction) - [Claude: yes — Anthropic Claude models are named in the SDK's supported-models table and used in its examples.](https://docs.cartesia.ai/line/sdk/agents) - [Open models: partial — The LLM layer routes through LiteLLM to 100+ providers, but the vendor names only Anthropic, OpenAI and Google, and its own speech models are API-only.](https://docs.cartesia.ai/line/sdk/agents) **Trade-offs against OpenAI Realtime API:** - [Multi-agent: no — System tools enumerate three slots: end call, DTMF, and transfer to an E.164 phone number. An agent therefore hands off to a number, not to another Cartesia agent.](https://docs.cartesia.ai/agents/system-tools) - [Agent permissions: partial — Limits are prompt-level guardrails plus per-tool execution settings (timeout, immediate vs async, speak-before-acting); there is no separate policy engine.](https://docs.cartesia.ai/agents/configuration) - [Admin controls: partial — Roles are Admin and Member only: an admin controls the org, invitations and members, not what an agent is permitted to do at runtime.](https://docs.cartesia.ai/enterprise/set-up-an-organization) [View Cartesia](https://theagentsindex.com/cartesia) ## Method Unknown data is never treated as a benefit or a trade-off. Featured placement is labelled but cannot alter recommendation order.