Inworld announced its Inworld Inference service on October 14, 2026, saying it is now open to every developer. The service targets consumer applications that make many model calls per interaction, such as conversational apps, tutors and games. It hosts selected open models and offers an OpenAI-compatible Chat Completions API.
Developers can select Flex, Default, Priority or Fast serving per request, API key or workspace. Inworld says Flex prioritizes lower cost and fails without billing if capacity is busy; Priority adds priority scheduling and model-dependent commitments. These are different serving choices, not different model weights, and availability depends on the model and service.
The system reuses matching prompt-prefix attention data and routes requests toward servers with warm caches while considering load. Inworld reports up to 90 percent cache reuse on named repeated-prompt workloads. That is not a promise that 90 percent of every customer's input will be cached; request similarity and placement matter.
The launch lists Gemma 4 26B Default pricing at USD 0.07 per million input tokens, USD 0.035 per million cached input tokens and USD 0.34 per million output tokens. Those are the vendor's named model and tier prices at launch, not a flat subscription or the price of every hosted model.
Inworld reports a Gemma 4 26B p95 first-token time of 0.31 seconds compared with 2.42 seconds for the median compared OpenRouter provider. Its source note says these are snapshots from each provider, not a matched-workload benchmark. The figures therefore should not be read as independently proven performance under every application's load.
Automatic fallback can try compatible models if the first model exceeds a configured first-token timeout. Response metadata records attempts and routing reasons. Inworld's GLM 5.3 Flash reliability claim covers more than 20 million requests over 30 days through October 12, with 98 failures; it is a vendor observation of that model and window, not a universal service guarantee.