Meta's Muse Spark 1.3 lands on OpenRouter with long-term agent focus

Muse Spark 1.3 became available on OpenRouter yesterday as a multi-modal reasoning model that Meta built specifically for agentic workflows involving multiple agents and code that require sustained execution. The model can track information accumulated over a long task, handle contradictory inputs that arise during operation, and ask the user for clarification or approval when it detects ambiguity, with the stated emphasis on concise execution without superfluous text.

Audio understanding in version 1.3 is not fully supported, and Meta explicitly notes that response quality may suffer on requests that include voice content. This limitation should be considered before building features that combine audio; the model is currently aimed at text and vision only.

The model offers a context window of one million tokens at a price of $1.25 per million input tokens and $4.25 per million output tokens. These figures place it in the mid-range of large reasoning models, cheaper than the premium offerings from OpenAI and Anthropic, but more expensive than open models that run locally.

In OpenRouter’s P50 benchmark the model delivers 81 tokens per second with a median retrieval latency of 3.23 seconds on the fastest provider. These numbers represent the best performance observed across providers running the model on the platform; actual figures may vary depending on the routing path chosen.

OpenRouter reports availability of 99.97% and 84.83% (two separate metrics the platform publishes) and enables automatic routing to a healthy provider when an error occurs with the primary provider, provided the request settings permit it. Per-provider availability data are accessible through the Endpoints API for anyone who wants to build custom fallback logic.