Meta Model API

Meta Model API is a hosted artificial-intelligence inference service from Meta Superintelligence Labs. Meta opened it in public preview on July 9, 2026, with Muse Spark 1.1 as its initial model. At launch, the service offered self-serve signup to developers in the United States and gave new accounts $20 in launch credits.^[1][2] The service has since added Muse Spark 1.2 (August 5, 2026, with what Meta called "expanded global access") and Muse Spark 1.3 (September 2, 2026), whose max reasoning mode followed on September 4, while Muse Spark list prices stayed at the July levels.^{[9][12][13][16]}

On October 12, 2026, Meta added Muse Wonder to the API. Muse Wonder Full launched through Muse Code subscriptions, not the API.^[21][22]

The API is distinct from both the consumer Meta AI service and Meta's open-weight Llama program. Muse Spark 1.1 also powers Thinking mode in Meta AI, but the Model API provides developer-controlled prompts, tools, and agent scaffolds through a paid endpoint. Meta's earlier Llama API preview served Llama models and offered different experimental capabilities.^[1][8]

Release and access

Meta announced global general availability in its September 24, 2026, Connect recap.^[24] Its product landing page nevertheless still labeled the service a public preview on October 13, an inconsistency in Meta's status descriptions.^[28] Access remains subject to geographic restrictions rather than availability in every country.^[25] Muse Wonder itself is explicitly in preview.^[22]

Meta's developer landing page tells users to point an existing OpenAI-compatible client at the service. Current integration documentation identifies POST /v1/responses as the primary endpoint, with /v1/chat/completions and an Anthropic-compatible /v1/messages interface also documented. The Responses surface has the broadest advertised feature support, including built-in web search.^[2][6][7]

Meta's detailed reference and contractual documents required sign-in during the launch period; its current documentation and service terms are publicly readable.^[26][27] No separate first-party Model API SDK or public source repository was identified at launch. Compatibility with the OpenAI API format lets developers reuse clients, but it does not imply that every OpenAI request field or behavior is supported.^[2][4][6]

Model roster after launch

Meta added Muse Spark 1.2 to the API on August 5, 2026, together with the Muse Code terminal agent, and said the model was available "in Muse Code and in Meta Model API with expanded global access" after the US-only July preview.^[9] The same release introduced a second, cheaper route: a Contributor tier at $0.10 per million input tokens and $0.20 per million output tokens for developers who allow Meta to use their prompts and completions to train its models, which MacRumors reported alongside the unchanged Standard prices (CNBC quoted Wang calling the tier "more than 10 times cheaper" than pay-as-you-go). MacRumors gave the Contributor rate limits as 60 requests and 2,100,000 tokens per minute against 3,000 requests and 4,000,000 tokens per minute on the Standard tier, and CNBC reported that Meta had begun accepting requests for zero-data retention.^[10][11]

Muse Spark 1.3 followed on September 2, 2026, available the same day through Muse Code and the API.^[12] Winbuzzer and TechRepublic both reported that the standard rates did not move: $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens, the same as for 1.2. Winbuzzer added that the model has a one-million-token combined context window and accepts text, images and video as input while producing text.^[14][15] Meta's own developer pages, as archived on September 3 and 4, list muse-spark-1.3 at those prices and muse-spark-1.3-contributor at $0.10 input, $0.002 cached input and $0.20 output, each with a 1M context window, next to identical rows for the two 1.2 variants.^[13][20]

At launch Meta withheld the 1.3 model's max reasoning mode pending additional safety testing; the launch post said "Previously available reasoning modes are available today with max reasoning coming shortly after we finish additional safety testing."^[17] On September 4, 2026, Alexandr Wang posted that Meta had "publicly released Muse Spark 1.3 max," and the Meta for Developers account wrote that "Muse Spark 1.3 with max reasoning is now available on Muse Code and Meta Model API."^[16][18] Meta then changed the launch post's availability line to that wording without an edit note.^[12][17]

The Contributor discount drew its own coverage after the 1.3 release. Cryptopolitan, citing Meta's pricing page as viewed on September 3, described the tier as "up to 21x" cheaper (the $0.20 output rate against $4.25) and its cached-input rate as a 75x cut, quoted Meta's pricing guide as saying the tier "lowers the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable," and gave the Contributor rate limit as 100 requests per minute against 3,000 on the standard plan, a higher Contributor figure than the 60 requests per minute MacRumors reported at the August launch.^[11][19]

Muse Wonder API preview

The API serves muse-wonder-1.0 in Standard mode only. Muse Wonder Full requires a Muse Code Power Usage or Maximum Usage subscription; selecting max reasoning on the API model does not select Full.^[22]

On Responses, the effort setting is reasoning.effort; on Chat Completions it is reasoning_effort. none is unsupported. Private reasoning consumes the output-token budget and is billed at the output-token rate.^[29]

The hosted catalog also includes Muse Image for image generation and editing, Muse Voice Transcribe for speech-to-text, and SAM 3.1 for segmentation. Muse Glimmer instead runs locally from open weights, not through Model API.^[26]

Model and modalities

Muse Spark 1.1 is a proprietary multimodal reasoning model designed for coding, tool use, computer interaction, and long, multi-step tasks. Meta advertises a one-million-token context window with active context management that can compact an ongoing task as it approaches the limit. The July launch materials did not specify a maximum output-token setting.^[1]

The launch materials describe understanding of text, images, video, PDF files, and audio. The direct API's output is text: Meta's evaluation report describes a query-only interface returning text completions, without exposing hidden chain-of-thought, gradients, or model activations.^[1][3][5]

The July integration documentation marked audio unsupported in its direct Model API integration.^[6] Meta's current documentation lists audio input for Muse Wonder and Spark, but warns that Spark 1.3 audio understanding is not fully supported and may produce degraded responses; it recommends Spark 1.2 or Muse Voice Transcribe for audio.^[22] The July launch announced no media-generation endpoint; the current catalog separately includes Muse Image.^[1][26]

Meta has not disclosed Muse Spark 1.1's parameter count, architecture, training corpus, training compute, or downloadable checkpoint. It has also not announced Model API fine-tuning, checkpoint export, or self-hosting. These omissions distinguish the service from Meta's open-weight Llama releases.^[1][8]

Tools and agent workflows

The API supports developer-defined function calling, developer prompts, and agent scaffolds. Meta says Muse Spark can generalize to native tools, custom skills, and Model Context Protocol servers. It can operate as a main agent that creates subagents and can coordinate parallel work when the developer's harness supplies those facilities.^[1][3]

Built-in web search can ground an answer with citations. Current integration documentation also lists JSON-schema structured output, prompt caching, reasoning-effort controls, and seed values. Their request fields are now documented in Meta's public guides.^[26]

Meta presents coding and computer use as important use cases. This does not give the model unrestricted access to a machine. A developer must provide a browser, shell, file system, application, or other tool and decide what the model may call. Similarly, MCP compatibility means the model can work with developer-connected servers; Meta is not hosting every MCP server on the user's behalf.^[1][4]

Meta's multi-agent examples also describe a system pattern rather than a guaranteed server-side workflow. The model can propose and coordinate subagents, but the surrounding application remains responsible for running them, enforcing permissions, handling failures, and accounting for their token and tool costs.^[1][4]

Pricing and rate limits

Reuters reported launch pricing of $1.25 per million input tokens and $4.25 per million output tokens. July integration documentation matched those prices and additionally listed cached input at $0.15 per million tokens and built-in web search at $2.50 per 1,000 queries.^[5][6]

July integration documentation listed per-team limits of 60 requests per minute and 2 million tokens per minute for initial-credit accounts, and 3,000 requests per minute and 4 million tokens per minute for paid use.^[6] The $20 launch credit was not described as a permanent free tier.^[2]

Meta's public pricing documentation listed the following Muse Wonder rates on October 13, 2026. It does not offer Wonder Contributor pricing on the API.^[23]

The same documentation lists per-team, per-model limits and charges web search at $2.50 per 1,000 queries in addition to tokens.^[23]

Meta now publishes cached-token, search and rate-limit figures directly, including the unchanged Muse Spark token rates. Account billing remains governed by the applicable service terms and rates.^[23][27]

Data handling and privacy

At the July 2026 launch, the accessible Meta pages did not state whether routine API prompts and responses were retained, used to train models, used for advertising, stored in a particular region, or deleted after a fixed period. The current service terms therefore matter for applications handling confidential, personal, or regulated information.^[27]

Since the August 2026 release of the Contributor tier, Meta's developer pages have drawn the training-use distinction explicitly: the standard muse-spark models are labeled "Not used to improve our products" and the -contributor variants "Used to improve our products."^[13][20] Current public terms prohibit training Meta models on Standard-request content, while Contributor content may be used for training. Contributor users must not submit sensitive, confidential or personal information. Retention depends on providing the service, legal obligations, permitted uses and safety or policy review; the terms do not promise a fixed deletion period.^[27]

This gap should not be filled with the earlier Llama API policy. In 2025, Meta explicitly said it would not use prompts or model responses from that limited preview to train models. No accessible Model API source identified for the July 2026 launch restates that commitment, so it cannot be assumed to carry over.^[8]

Meta's Muse Spark evaluation report does disclose a narrower safety practice: safeguards may aggregate and analyze longer-term activity patterns to identify persistent malicious use. That statement concerns abuse detection and is not a complete retention, training, or privacy policy.^[3]

Safety and acceptable use

Meta evaluated Muse Spark 1.1 under its Advanced AI Scaling Framework in Chemical and Biological, Cybersecurity, and Loss of Control risk domains. The report describes unmitigated high or possible-high capability findings in some tests. Meta says layered mitigations reduce the remaining risk to moderate or lower under its own framework.^[3]

The deployed system combines refusal behavior, prompt-injection testing, abuse monitoring, and system-level safeguards. Meta reports testing direct jailbreaks, attacks hidden in untrusted tool data, and attempts to override developer instructions. The report refers to an API content-safety policy, but the complete current acceptable-use terms are not reproduced in the public report.^[1][3]

Meta also assigns responsibility to application developers. Its report recommends strict tool allowlists, isolated workspaces, and system-level controls suited to the deployment. These controls are especially important for AI agents that can execute code or act on external services, because a capable model-level defense does not eliminate application permissions, data-exfiltration, or prompt-injection risks.^[3]

The safety findings are Meta's pre-deployment assessment, not an independent certification. The report exposes useful methods and results, but it does not show that every third-party agent configuration is safe or that all harmful uses will be detected.^[3]

Relationship to Meta AI and Llama API

Muse Spark 1.1 is the model, while Meta Model API is the hosted developer service. The same model appears in Thinking mode in Meta AI, where Meta controls the consumer interface and system scaffolding. A developer using the API supplies a different prompt, tool set, permission model, and application context, which can change both capability and risk.^[1][3]

The Model API is also not simply a new name for the Llama API. Meta announced the Llama API in 2025 as a limited preview for Llama 4 Scout and Maverick, with Python and TypeScript SDKs, evaluation and fine-tuning experiments, and a path to export tuned weights. Current Promptfoo documentation says that preview was retired on July 6, 2026, three days before the Model API launch.^[6][8]

The replacement-era service began with a closed Muse model and usage-based pricing. It offers no public weight export or fine-tuning path. Developers seeking open-weight deployment therefore still need the Llama ecosystem or another model host, while those using Meta Model API depend on Meta's endpoint, model availability, and evolving service terms.^[1][2][6]

Limitations

Meta Model API began with a narrow launch region and one model, then expanded its model roster and access.^[1][9][13] Meta announced general availability in September, although its product page retains preview wording; Muse Wonder remains a preview model.^[22][24][28] Geographic restrictions and the applicable service terms still limit deployments.^[25][27]

The published one-million-token context window does not guarantee that every request can use the full amount with every tool or attachment, and active compaction may summarize earlier material. Likewise, model support for MCP, computer use, and subagents depends on developer-supplied infrastructure. These are capabilities for building an agent system, not a hosted promise that the API will safely operate arbitrary software by itself.^[1][4]

Muse Spark 1.1's weights, architecture, training data, and parameter count remain undisclosed. Meta's benchmark and safety results are vendor-reported. Audio support has model-specific limits, including the Spark 1.3 caveat.^[22] Compatibility, policies, limits and the model roster remain subject to change; Meta's service terms reserve the right to alter functionality and access.^[27]

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Meta Superintelligence Labs. Introducing Muse Spark 1.1. Meta AI, July 9, 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6Meta. Muse Spark on Meta Model API. Meta for Developers. Accessed July 24, 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Meta Superintelligence Labs. Muse Spark 1.1 Evaluation Report. July 9, 2026.
4. ^1 ^2 ^3 ^4Meta. Build with Muse Spark, now available on Meta Model API. Meta for Developers, July 2026.
5. ^1 ^2Mukherjee, Supantha. Meta debuts Muse Spark 1.1 with preview open to developers. Reuters, July 9, 2026.
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7Promptfoo. Meta Model API provider. Updated July 18, 2026.
7. ^OpenClaw. Meta Model API. Accessed July 24, 2026.
8. ^1 ^2 ^3 ^4Meta. The Llama Ecosystem: Past, Present, and Future. Meta AI, April 29, 2025.
9. ^1 ^2 ^3 ^4Meta Superintelligence Labs. Introducing Muse Code and Muse Spark 1.2. Meta AI Research, August 5, 2026.
10. ^Vanian, Jonathan. Meta debuts first AI coding agent to take on Anthropic and OpenAI. CNBC, August 5, 2026.
11. ^1 ^2Clover, Juli. Meta's New Mac Coding Agent Costs Up to 20x Less If You Let Meta Train on Your Data. MacRumors, August 5, 2026.
12. ^1 ^2 ^3Meta Superintelligence Labs. Introducing Muse Spark 1.3. Meta AI Research, September 2, 2026; availability line updated in place by September 5, 2026.
13. ^1 ^2 ^3 ^4 ^5Meta. Muse Spark 1.3. Meta for Developers, as archived by the Internet Archive on September 3, 2026.
14. ^Kasanmascheff, Markus. Meta Releases Muse Spark 1.3 Model for Longer Tool-Based Work. Winbuzzer, September 4, 2026.
15. ^Abdullahi, Aminu. Meta Says Muse Spark 1.3 Beats GPT-5.6 Sol at Coding: Independent Tests Are More Mixed. TechRepublic, September 4, 2026.
16. ^1 ^2Wang, Alexandr (@alexandr_wang). 1/ we just publicly released Muse Spark 1.3 max!. X, September 4, 2026.
17. ^1 ^2Meta Superintelligence Labs. Introducing Muse Spark 1.3, Internet Archive snapshot of September 2, 2026 (19:37 UTC). Meta AI Research, September 2, 2026.
18. ^Meta for Developers (@MetaforDevs). Muse Spark 1.3 with max reasoning is now available on Muse Code and Meta Model API. X, September 4, 2026.
19. ^Moses, Randa. Meta cuts Muse Spark prices up to 21x for developers who hand over their prompts. Cryptopolitan, September 3, 2026.
20. ^1 ^2Meta. Muse Code. Meta for Developers, as archived by the Internet Archive on September 4, 2026.
21. ^1 ^2Meta Superintelligence Labs. Muse Wonder: A Frontier Model For Personal & Professional AI Agents. October 12, 2026.
22. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Meta. Models. Meta Model API documentation. Accessed October 13, 2026.
23. ^1 ^2 ^3Meta. Pricing and rate limits. Meta Model API documentation. Accessed October 13, 2026.
24. ^1 ^2 ^3Meta. Meta Connect 2026: The end-to-end recap. September 24, 2026.
25. ^1 ^2 ^3Meta. Meta Model API Model Designation and Geographic Use Policy. Updated October 12, 2026.
26. ^1 ^2 ^3 ^4Meta. Build with Meta Model API and Muse Code. Meta Model API documentation. Accessed October 13, 2026.
27. ^1 ^2 ^3 ^4 ^5 ^6Meta. Meta Model API Terms of Service. Updated October 9, 2026.
28. ^1 ^2 ^3 ^4Meta. Meta Model API. Product overview. Accessed October 13, 2026.
29. ^1 ^2Meta. Reasoning. Meta Model API documentation. Accessed October 13, 2026.