Model, application or API?
A hosted model produces a response. A coding agent supplies the surrounding tools and runs a workflow. A consumer application supplies its own interface and account experience. Buying access to one route does not automatically purchase the others.
| Your task | Starting point | Budget question |
|---|---|---|
| Build your own text-based app or agent | Meta Model API | Tokens, tools and requests |
| Use an agent in a repository | Muse Code | Subscription allowance or usage billing |
| Create or edit pictures | Muse Image | Successfully generated images |
| Run weights yourself | Spark Open or Glimmer | Infrastructure, operations and license |
| Use the consumer assistant | Official app route | Your account’s current offer |
Wonder and Spark
Meta positions Wonder for demanding, extended agent work and Spark for cost-sensitive agent workflows. Both accept multiple input modalities and return text. This describes the provider’s intended use, not a measured success rate on your codebase.
Wonder’s hosted ID is muse-wonder-1.0. Spark’s versioned hosted IDs include muse-spark-1.3, muse-spark-1.2 and muse-spark-1.1. Use explicit IDs in a rollout log so that a later comparison measures the same model.
Wonder Full is a separate subscription-access variant. It is not the result of raising regular Wonder’s reasoning effort to max. See Full versus max effort before comparing an API bill with a subscription offer.
Input understanding is different from output generation
Sending an image or video to a text-output model asks it to understand that input. To receive a generated image, use Muse Image. The fact that a model accepts video does not make it a video-generation service. Keep generation and understanding separate in your feature specification and budget.
For an image workflow, decide whether the desired output is a textual description, an edited picture or segmentation geometry. Those are different API tasks; pricing a text response as if it were a generated image produces the wrong estimate.
A practical selection workflow
- Write down the expected output and the tools your application must call.
- Choose a hosted API, a ready-made coding agent or self-hosted weights.
- Run a small representative evaluation with a fixed task set and spend cap.
- Measure accepted work, retries and total cost, rather than comparing input price alone.
- Record the model ID, reasoning effort, provider, client and date alongside results.
Use the same-workload cost comparison for budgeting. Use the evaluation guide when deciding whether additional model cost is justified.
Sources & evidence
Editorial check: 17 October 2026. Product facts come from the following primary sources; planning checklists and examples are our explanations. No paid inference, model benchmark or regional checkout was run for this guide.