Inworld Realtime API vs Pipecat PhoneLLM

Inworld Realtime API 71

Release note of 16 October 2026 (cache_transcript), 21 days ago (30). Three dated Realtime entries in the 90 days from 8 August, on 14 September, 23 September and 16 October (20). Discord and support@inworld.ai, with no evidence of answers read (8 of 15). Works with OpenAI's Realtime clients and the CLI is current at 1.5.0, but there is no Inworld Realtime SDK (8 of 15). CLI on npm under MIT, CI not read (5 of 10). Total 71.

Pipecat PhoneLLM 58

Model reading, with notice periods and churn in place of release counts. Alpha 2 released on 15 October 2026, and no newer PhoneLLM repository or file change is on Hugging Face on 4 November (30). Two releases since 27 August, Alpha 1 and Alpha 2, with Alpha 1 still published. Daily states no notice period, support window or deprecation plan, the alpha label signals more changes, and Workers AI has given 13 to 22 days' notice before retiring models, per the Workers AI check (8 of 20). Three Alpha 1 discussions opened on 31 August and 16 September, a vLLM caching bug, a request for FP8 weights and a tool-call fault, still have no reply on 4 November (one post each, counted by script), though FP8 weights did ship with Alpha 2. Alpha 2's two repositories have no discussions. Daily's Discord is linked from the post and wasn't sampled (4 of 25). vLLM 0.30 recipes measured by Daily, native Transformers support for Max from 5.8, and Workers AI with Cloudflare's SDKs. Mini still needs repository code (12 of 15). Max's generation config names Transformers 5.8.1 and Mini's 4.57.6. No CI or tests for the weights visible to us (4 of 10). 58.

Inworld Realtime API 60

No AsyncAPI or other machine-readable file for the socket, and the spec paths redirect (0). llms.txt, Markdown twins and a docs MCP server (10). The overview, OpenAI migration guide, providerData reference and the billing page's account of how a session is counted say what to use when (16 of 20). The reference lists every client and server event with field types and required marks, as text rather than a schema (12 of 15). JavaScript and Python example projects and an error event with type, code and param, without a list of codes (9 of 15). Dated release notes with affected clients and required actions, but no API version beyond /v1 (13 of 15). Total 60.

Pipecat PhoneLLM 72

Model reading on the API lines. Each Workers AI model page links JSON Schemas for input and output. Mini's synchronous input schema lists required: ["prompt"] and has no messages property, while the page's TypeScript, Python and curl examples and Cloudflare's changelog all send messages. It also carries audio, modalities and web_search_options, which a text-only model has no use for (18 of 25). developers.cloudflare.com/llms.txt and a Markdown copy of the model page and changelog. The cards are Markdown on Hugging Face (10). The post and cards say what each size is for (Mini for speed, Max for accuracy, both for phone agents with tool calling and Pipecat Flows). Neither card has a limitations or out-of-scope section, and English-only is in the metadata, not the text (12 of 20). temperature bounded 0 to 2, tool_choice an enum or a forced function, response_format a typed one-of, and chat_template_kwargs.enable_thinking typed and described (12 of 15). Examples in TypeScript, Python, curl and vLLM recipes for both sizes. The Workers AI examples omit the temperature: 0 and enable_thinking: false that Cloudflare's changelog asks for, and the error table is Workers AI's generic one (11 of 15). The version is in the model ID (alpha-2), Alpha 1 stays published, and Hugging Face keeps commit history (three commits each). No changelog of PhoneLLM's own. Changes are listed in the post and Cloudflare's changelog entry (9 of 15). 72.

Inworld Realtime API 58

Workspace API keys sent as Basic credentials, deleted to revoke, with separate keys per environment advised. The CLI mints read-only keys unless --write is given, but write toggles cover only the Voice and Router APIs, so no key is limited to one product. Browsers use short-lived tokens minted server side. Session IDs, not secrets, go in the URL (25 of 30). Tools run in the client, so the developer approves side effects, with no approval step in the API (10 of 20). The agent listens to callers and passes their speech to an LLM, and no prompt-injection guidance was found (5 of 15). Usage per key and a Performance page with errors and throttling (8 of 15). SOC 2 Type II and HIPAA claimed and a vulnerability report form, with no security.txt or bug bounty, per the 4 November reading (10 of 20). Total 58. Zero data retention covers session content from the Developer plan up, but not requests routed to third-party LLM providers.

Pipecat PhoneLLM 69

Model reading. On Workers AI, Cloudflare API tokens limited to Workers AI permissions on chosen accounts, revocable, with expiry and IP filters, per the Workers AI listing's checks. Permission is per account, not per model (27 of 30). Cloudflare's data usage page and service terms say customer content isn't used to train models without consent, with clause 9 allowing use to run and improve services, per the same checks. The model page now marks Zero data retention as Yes, which we haven't matched to a definition page, so the line holds. Self-hosted, prompts stay on the owner's hardware (17 of 20). The model is built to take a caller's transcribed speech and call tools. Neither card nor the post has limitations, a safety evaluation, or guidance on injected or hostile input. The template wraps tool results in <tool_response> and does nothing to neutralise caller text. Mini self-hosted runs modelling code from the repository with --trust-remote-code (3 of 15). Direct Workers AI calls have no per-request log that we found. AI Gateway logs prompts and responses when a call is routed through it (6 of 15). For the hosted route, Cloudflare's security.txt names HackerOne, and its ISO page lists ISO 27001:2022, 27018 and 27701 for the global cloud platform, which the Cloudflare listings read at 18 in this run. Daily has no security.txt and a disclosure policy with no bounty, per the Pipecat check, and the cards give no security contact, so 2 held back from Cloudflare's value (16 of 20). 27 + 17 + 3 + 6 + 16 = 69.

Inworld Realtime API 68

Scored as an API. Conversation history defaults to the last 100 items, can be trimmed, and can be cached on Anthropic models, and every response.done carries a usage block (18 of 25). No lists to page. Output controls are modalities, turn detection settings, optional transcription and response.speak (12 of 20). Error events follow OpenAI's type and code shape, unsupported fields come back as non-fatal unsupported_field warnings, and item_not_found names the bad parameter (14 of 20). Tools run on the client and a follow-up response starts automatically after a tool result. Nothing covers side effects of a tool call cut off by an interruption (12 of 20). Works with OpenAI Realtime clients after changing the URL and key, with no Inworld Realtime SDK (12 of 15). Total 68.

Pipecat PhoneLLM 61

Model reading (tool use, structured output, caching, context, batch, SDKs, errors and defaults). Trained for multi-turn tool calls and Pipecat Flows, with OpenAI-style tools, tool_choice and parallel_tool_calls on Workers AI. Daily's PhoneBench gives 5.9% tool-call failures for Mini and 3.1% for Max (vendor-reported). Alpha 1 discussion 3 reports the model confirming details back in place of calling a tool, unanswered, and we don't know whether Alpha 2 fixes it (15 of 20). json_object and json_schema in the Workers AI schema. The Workers AI docs say adherence isn't guaranteed, and neither card mentions structured output (8 of 15). The vLLM recipes turn on prefix caching, but Alpha 1 discussion 1 reports it as a no-op for nemotron_h in vLLM, unanswered. Workers AI lists no cached price for Mini (5 of 15). 262,140 tokens on Workers AI and 262,144 in Mini's config, 1,048,576 in Max's config. The cards' recipes cap at 32,768 (13 of 15). The Mini model page has no Batch tag, so the Workers AI batch route isn't open to it. Self-hosted, the cards suggest 128 to 256 requests in flight per GPU (3 of 10). Any OpenAI SDK against /v1/chat/completions, the Workers binding, Cloudflare's TypeScript, Python and Go SDKs, and Pipecat's OpenAI-compatible services (9 of 10). Workers AI errors carry internal codes. Defaults work against the caller. Both repositories' generation_config.json sample by default (Max at temperature 1.0) where Daily recommends 0, and Max needs chat_template_kwargs on every request or its reply lands in reasoning (8 of 15). 61.

Inworld Realtime API 63

Closed service with clear terms naming Theai, Inc. dba Inworld AI (15). The ZDR page says what a session keeps and that third-party LLM requests aren't covered, but the terms and privacy notice disagree on training (12 of 30). Release notes flag potentially breaking changes with the clients affected and what to do, without stating a notice period (14 of 20). US headquarters with EU and India residency on Enterprise, and no sub-processor list in the docs, while sessions may route to third-party LLM providers (10 of 20). Total 51.

Pipecat PhoneLLM 58

Mini is BSD-2-Clause, an OSI licence, with NVIDIA's Nemotron Open Model License kept for the base work, and the card explains how the two apply. Max is OpenMDW-1.1, permissive but not on the OSI list. Training data and code aren't published, and the hosted route runs under Cloudflare's terms (22 of 30). The post says training data is fully synthetic, from unnamed open-weight teachers. On Workers AI, Cloudflare's no-training statements agree with each other apart from clause 9, per the Workers AI check. Self-hosted, nothing leaves the owner's hardware. Daily's privacy policy and DPA weren't re-read for the models (18 of 30). No deprecation policy or dated notices from Daily. Workers AI posts dated notices but has no policy (5 of 20). Cloudflare publishes a sub-processor page, and where a Workers AI request runs isn't stated beyond its network. The weights are safetensors, but Mini's repository ships Python modelling code that a self-hosted run executes (10 of 20). 55.