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. The service is available through self-serve signup to developers in the United States and gives new accounts $20 in launch credits.^[1][2]

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 describes the service as a public preview rather than a generally available product. Its public developer page limits eligibility to US developers, offers one-click API-key creation, and promotes compatibility with existing OpenAI SDK code. Meta has not announced a general-availability date, international rollout schedule, enterprise service-level agreement, or dedicated-capacity option on its accessible public pages.^[2]

Scroll sideways for more →

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 has published quick-start and agent-pattern resources, but its detailed API reference and current contractual documents redirect to a signed-in developer portal. 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 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. Meta has not published a maximum output-token setting for the public preview.^[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]

Audio availability needs qualification. Meta says Muse Spark can inspect audio, but current Promptfoo provider documentation marks both audio input and output unsupported in its direct Model API integration. Consequently, the public evidence establishes broad model-level audio capability, not reliable audio acceptance on the present API surface. No image-, video-, or audio-generation endpoint was announced.^[1][6]

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. These are preview features whose exact request fields and limits are governed by the signed-in documentation.^[2][6]

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. Current Promptfoo documentation matches those prices and additionally lists cached input at $0.15 per million tokens and built-in web search at $2.50 per 1,000 queries.^[5][6]

Scroll sideways for more →

Promptfoo lists per-team limits of 60 requests per minute and 2 million tokens per minute for accounts using the initial credit, and 3,000 requests per minute and 4 million tokens per minute for paid use. Meta's public landing page confirms the $20 credit but does not characterize it as a permanent free tier. It also does not publicly list batch discounts or a capacity reservation program.^[2][6]

The cached-token, search, and rate-limit figures come from current open-source integration documentation rather than an accessible public Meta pricing contract. They may change during the preview, so production users need to confirm the values shown in their Meta developer account.^[6]

Data handling and privacy

The accessible Meta launch pages do not state whether routine API prompts and responses are retained, used to train models, used for advertising, stored in a particular region, or deleted after a fixed period. The signed-in terms therefore matter for applications handling confidential, personal, or regulated information.

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 begins 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, preview availability, and evolving service terms.^[1][2][6]

Limitations

As a public preview, Meta Model API has a narrow launch region, one publicly documented model, no announced stability guarantee, and no public general-availability schedule. The detailed reference, terms, and account-specific limits sit behind sign-in, leaving important data-handling and enterprise-operational questions unanswered on the open web.^[2][6]

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, and present audio support is unclear across public sources. The service is therefore usable as a documented preview, but its compatibility, policies, limits, and model roster should be treated as changeable until Meta publishes a stable general-availability contract.^[1][2][3][6]

References

1. Meta Superintelligence Labs. Introducing Muse Spark 1.1. Meta AI, July 9, 2026. ↩
2. Meta. Muse Spark on Meta Model API. Meta for Developers. Accessed July 24, 2026. ↩
3. Meta Superintelligence Labs. Muse Spark 1.1 Evaluation Report. July 9, 2026. ↩
4. Meta. Build with Muse Spark, now available on Meta Model API. Meta for Developers, July 2026. ↩
5. Mukherjee, Supantha. Meta debuts Muse Spark 1.1 with preview open to developers. Reuters, July 9, 2026. ↩
6. Promptfoo. Meta Model API provider. Updated July 18, 2026. ↩
7. OpenClaw. Meta Model API. Accessed July 24, 2026. ↩
8. Meta. The Llama Ecosystem: Past, Present, and Future. Meta AI, April 29, 2025. ↩