Inworld Router, now Inworld Inference: pricing and limits

This page is built from Inworld AI's inference, models, pricing, security and terms pages, its docs and llms.txt, its public status tracker, archived copies of the old router page, Inworld's own launch measurements and comparison page, a competitor's review and public user reviews, all checked on October 21, 2026. Inworld's claims are labelled as its own.

What changed on October 14, 2026

Inworld AI sold Inworld Router, also called LLM Router and Realtime Router, as one OpenAI- and Anthropic-compatible endpoint for third-party models. Archived copies of inworld.ai/router promised "One API for 200+ AI Models" in March 2026 and "220+ models, zero markup" on October 4, 2026. On October 14, 2026 Inworld launched Inworld Inference, which serves open models such as Gemma 4, GLM, DeepSeek and MiMo on Inworld's own hardware, and inworld.ai/router now redirects to inworld.ai/inference.

The routing features did not disappear. The docs keep them under the Inference name, at the same /router/ paths: saved routers, conditional routes, traffic splits, provider fallbacks and request-level routing. The launch post does not mention the Router name or say what happens to third-party routing. Inworld's models page listed 129 models on October 21 (109 of them text), against the 220+ claimed on October 4; Inworld gives no reason for the difference.

What Inworld Router sells

You send OpenAI-style chat completions or Responses requests to api.inworld.ai/v1, or Anthropic-style messages to api.inworld.ai, and name a model (anthropic/claude-opus-5-5, openai/gpt-6-sol) or a saved router (inworld/<router-id>). Providers named in the docs include OpenAI, Anthropic, Google AI Studio, Vertex AI, DeepInfra, DeepSeek, Mistral, Groq and xAI. A saved router can choose a model from request metadata with CEL expressions (language, plan, emotion, message length), split traffic by weight with sticky user assignment for A/B tests, and export exposure data to BigQuery or Mixpanel. The same key and balance cover image and video generation and Inworld's speech APIs; a chat completion can even return synthesized speech.

Prices against the official APIs

Inworld's billing docs say: "For LLM requests routed to third-party providers through Inworld, you pay the provider's listed cost with no markup." For the models checked, its models page matched the makers:

Inworld's prices come from its models page. Some things cost extra: each web search step adds the search engine's fee, every model in a fallback chain that produced output is billed, and a BYOK request whose key the provider rejects is retried on Inworld's key and billed as Inworld usage. Plans from Creator (25 USD a month) to Growth (1,500 USD) return their price as credits; their discounts apply to speech, not LLMs. The minimum credit purchase is 10 USD, extra credits expire after one year, and the terms make payments non-refundable except where law requires.

Inworld's own hosted models have their own prices. Inworld says Gemma 4 26B, at 0.07 USD per million input tokens, is nearly 40% below the median of other hosts' list prices on October 12, 2026.

What public tests show

No third-party benchmark of Inworld Router has been published. The numbers available are Inworld's own:

• Inworld's launch post (October 14, 2026) reports Gemma 4 26B at a p95 time to first token of 0.31 s against 2.42 s for the median OpenRouter provider, and GLM 5.3 Flash completing 99.999% of more than 20 million requests over 30 days. Inworld says its figures were measured at its own router while the OpenRouter figures are provider snapshots, not a matched benchmark. They cover models Inworld hosts, not third-party routing.
• Inworld's comparison page with OpenRouter claims zero markup and CEL routing and sticky A/B splits that OpenRouter lacks. The page warns that its latency headline is not a controlled benchmark and that Inworld has a commercial interest.
• Respan, a competitor, confirms in a desk review that there is no markup, including on BYOK traffic, and lists missing built-in evals, request-level observability only and hand-written CEL rules.

Reliability and user reports

Inworld's status page runs on a public GitHub issue tracker. Router incidents include intermittent LLM Router outages on April 12, 2026 caused by a Mistral incident (about 1.5 hours), elevated Realtime Router errors on May 16 (about 1 hour 40 minutes) and elevated latency on DeepSeek and Gemma models on July 16 (about 1 hour 20 minutes). Inworld publishes no uptime percentage.

User reviews are about Inworld's voice products: Product Hunt shows 5.0 from 7 reviews praising its text-to-speech. No user discussion of the router turned up on Hacker News or GitHub. Inworld's customer pages cite Status by Wishroll (AI costs down more than 95%) and Latitude (more than 25 million inference calls in six weeks); these are vendor case studies.

MCP, Agent Skills and coding agents

Inworld ships an official MCP server inside its CLI (@inworld/cli). inworld mcp runs over stdio, or over HTTP with --http, and exposes text-to-speech, voice design, speech-to-text, LLM chat, workspace limits, a billing summary and router management as tools. It runs as the signed-in CLI user; the docs show claude mcp add inworld -- inworld mcp for Claude Code. There is no hosted documentation MCP server; agents are pointed to llms.txt and Markdown pages instead.

No official Agent Skills were found in Inworld's docs or on its GitHub. Its Anthropic-compatible endpoint drops the mcp_servers and container fields, so Anthropic's MCP connector and Skills API do not work through it. It also drops PDF documents and thinking blocks, so reasoning across tool-call turns loses its thinking signatures, and it forwards Anthropic server tools such as web search as plain function definitions without running them.

The only coding agent guide is for Claude Code: set ANTHROPIC_BASE_URL to https://api.inworld.ai and ANTHROPIC_AUTH_TOKEN to your Inworld key, and Claude Code can also call models such as openai/gpt-5.4. The docs warn that the x-api-key header is not read, so a plain ANTHROPIC_API_KEY setup fails. There are migration guides from OpenRouter and Anthropic.

The OpenAI side supports chat completions, the Responses API with web search, tool calling (parallel_tool_calls and strict are ignored), structured outputs, reasoning effort, prompt caching and streaming. No LLM batch endpoint is documented, there is no Gemini-native format, and Inworld's SDKs cover speech; for LLMs you use the OpenAI or Anthropic SDKs.

Privacy, terms and who runs it

Theai, Inc., doing business as Inworld AI, is based in Mountain View, California. It was founded in 2021, has raised over 100 million USD, and bought the voice-agent startup Ultravox on October 1, 2026. Its security page claims SOC 2 Type II, HIPAA and GDPR compliance. Zero data retention covers speech and Inworld-hosted LLMs and can be switched only on Enterprise plans; requests routed to third-party providers fall under those providers' policies.

How it compares with OpenRouter, Vercel and Requesty

• OpenRouter adds 5.5% to credit purchases on its Standard plan but lists far more open-weight hosts and free models.
• Vercel AI Gateway also adds no markup, lists 482 models and documents setup for 29 coding agents.
• Requesty adds 5% to each top-up.

Who should skip it

Skip it if you want the broadest catalog, free models, a relay with an independent track record, or Anthropic features such as the MCP connector, Skills, PDFs or thinking across turns. The product has just been renamed and refocused on Inworld's hosted models, so check that the third-party models you need are still listed before building on it. It fits consumer and voice apps that already use Inworld's speech APIs and want per-user model routing and A/B tests at list prices.