Muse Spark Playground

Connection

Use https://api.meta.ai/v1 or your Microsoft Foundry OpenAI v1 endpoint. Foundry estimates use Meta prices, not published Azure rates.

Endpoint

API

Model (capabilities & pricing)

Authentication Mode

API Key

1,048,576-token context. Set a model independently of the deployment name so custom deployments use the right limits and price.

Advanced

Model controls

Maximum output tokens

Reasoning effort

Stream

Muse Spark always reasons. Reasoning and visible text share the output budget. Only Standard 1.3 supports max effort; its documented output limit is 131,072 tokens.

Reasoning summary

Temperature (0-2)

Top P (greater than 0, up to 1)

Frequency penalty (-2 to 2)

Presence penalty (-2 to 2)

Service tier

Sampling controls are supported. Tune temperature or Top P rather than both. Service tier is a compatibility setting with no published Meta SLA/price mapping; it does not select Contributor pricing.

Include stream-size obfuscation (API default)

Structured output

Output format

Developer tools & tool search

Enable developer-defined tools

Tool choice

Allow parallel function calls

Maximum built-in tool calls

Tool search

Hosted search needs at least one deferred tool. Client search pauses for a JSON array of discovered definitions. Custom tools are a documented compatibility surface with incomplete Meta result schemas; provide their provider-native result items manually.

Conversation, caching & identifiers

Store responses and use server-managed history

Previous response ID (optional starting point)

Include encrypted reasoning for client-managed replay

Background generation with automatic polling (storage on, streaming off)

The playground defaults to client-managed history (Meta Responses defaults to store:true). Encrypted content is opaque and never shown as a thinking transcript. Stored requests resend per-turn settings. Stopping a client connection may not stop server generation.

Prompt cache key

Prompt cache retention

Prefix caching is automatic, including with store:false. Retention is a hint, not a guarantee; keep stable instructions and shared prefixes first.

Safety identifier (up to 64 characters)

User identifier (deprecated)

Metadata (JSON object, up to 16 string pairs)

Advanced input & compatibility

Additional API-native input items (JSON array)

Appended after the composed user message. Use the selected API's exact message/content, role/name, tool-output, and replay formats. This is also the escape hatch for documented compatibility extensions; changing API does not translate these items.

Truncation

Request preview context compaction

Meta's guide and shared schema disagree on auto truncation, and compaction is preview-dependent. Both are off by default; the provider may reject or ignore them. Do not rely on them to fit the context window.

Unsupported or ineffective OpenAI controls are deliberately not sent: reasoning:none, logprobs, stop, verbosity, logit_bias, prediction, audio output/modalities, n>1, hosted prompt templates, and client_metadata. Muse Spark outputs text only. Legacy max_tokens and generate_summary aliases are replaced by the supported controls above.

Model	Input	Cached input	Output
muse-spark-1.3	$1.25	$0.15	$4.25
muse-spark-1.2	$1.25	$0.15	$4.25
muse-spark-1.1	$1.25	$0.15	$4.25
muse-spark-1.3-contributor	$0.10	$0.002	$0.20
muse-spark-1.2-contributor	$0.10	$0.002	$0.20

Model	Input	Cached input	Output
muse-spark-1.3	$1.25	$0.15	$4.25
muse-spark-1.2	$1.25	$0.15	$4.25
muse-spark-1.1	$1.25	$0.15	$4.25
muse-spark-1.3-contributor	$0.10	$0.002	$0.20
muse-spark-1.2-contributor	$0.10	$0.002	$0.20