On 9 July 2026, Meta launched the Meta paid AI API alongside Muse Spark 1.1 — the first time the company has charged for programmatic access to a frontier model. Pricing sits at $1.25/M input and $4.25/M output tokens, placing Meta at the lowest price floor among major US providers and undercutting OpenAI and Anthropic on output by 5-10x. The strategic story matters more than the model: the company that built its AI reputation on giving Llama away for free is now selling capacity at prices that would not sustain a pure-play lab.
TL;DR
- Last verified: 10 July 2026.
- Muse Spark 1.1 is a multimodal reasoning model from Meta Superintelligence Labs with a 1M-token context window.
- The Meta Model API (public preview, US only) charges $1.25/M input and $4.25/M output, cached input $0.15/M, web-search grounding $2.50/1K queries. New accounts get $20 in credits.
- The API accepts OpenAI-compatible and Anthropic-compatible request formats — migration cost is minimal.
- Consumers keep free access in "Thinking" mode in the Meta AI app and on meta.ai.
- Meta's shift from open-weights to a closed, paid frontier model is the real headline.
What is the Meta paid AI API and why does it matter?
The Meta paid AI API is Meta's first commercial developer endpoint for a proprietary frontier model, launched 9 July 2026. Previously, Meta released Llama weights openly and offered the original Muse Spark (April 2026) only via private preview. Muse Spark 1.1 is closed-weight and monetised per-token — a structural change in Meta's AI strategy.
The significance is competitive, not technical. OpenAI and Anthropic depend on API revenue to fund training runs. Meta does not. Its ad business generates enough cash flow that model access can be priced as a strategic instrument, pressuring labs that need higher margins.
How much does Muse Spark 1.1 cost compared with rivals?
Muse Spark 1.1 is the cheapest tier-one model on output tokens among major US vendors — cheaper than Anthropic's Opus line by an order of magnitude.
| Provider / Model | Input (\(/M tokens) | Output (\)/M tokens) | |
|---|---|---|
| Meta — Muse Spark 1.1 | 1.25 | 4.25 |
| xAI — Grok 4.5 (frontier) | ~2 | ~10 |
| OpenAI — GPT-5.5 | ~5–10 | ~15–30 |
| Anthropic — Opus 4.8 | ~10–15 | ~25–50 |
Cached input drops to $0.15/M — significant for agentic workloads that reuse long system prompts. Web-search grounding is $2.50/1K queries. For teams on OpenAI or Anthropic, the compatibility shims mean migration is largely a base URL and key change.
What can Muse Spark 1.1 actually do?
Muse Spark 1.1 is a multimodal reasoning model designed for agentic workflows. Its declared capabilities are:
- Modality: text, images, video, and PDFs as native inputs.
- Context: 1 million tokens with active compression for long-running agent sessions.
- Agent primitives: tool/function calling, structured outputs, computer use, and web-search grounding.
- Orchestration: works as a lead agent spawning parallel subagents, or as a subagent invoked by another orchestrator.
On vendor-reported benchmarks, Muse Spark 1.1 leads on MCP Atlas (88.1), Humanity's Last Exam (62.1), and JobBench. On SWE-Bench Pro it trails Opus 4.8 (61.5 vs 69.2), so heavy code-refactor workloads still favour Opus. Independent evaluator VALS-AI ranks it fourth overall, highlighting speed and cost-effectiveness. Vendor benchmarks are self-reported; validate against your own workload.
Is Muse Spark 1.1 free anywhere?
Yes, for consumers only. Muse Spark 1.1 is free in "Thinking" mode in the Meta AI app and on meta.ai — mirroring the pattern of other frontier labs: free for end users to defend consumer surface area, paid for developers through the API. Exploratory chat needs no billing. Products and backend services use the paid API.
Why did Meta abandon the open-source model for its frontier tier?
Meta has not abandoned open-source entirely, but has drawn a line at the frontier. Llama remains the open option for self-hosters. Muse Spark 1.1 sits above that line: closed weights, hosted-only inference, commercial licence via the API. The reasons: training economics of a trillion-parameter-class model, complexity of computer-use and web-grounding tools, and strategic control over how the model is used at scale. See our self-hosting AI models guide for what open-weights still offer.
How does this reshape the AI price war?
Competitive pressure now comes from a well-capitalised incumbent that does not need API margin — a new problem for pure-play labs. OpenAI and Anthropic can compete on capability, ecosystem, and enterprise trust, but cannot easily match Meta on price without eroding revenue that funds their next training run. xAI has already signalled aggressive Grok 4.5 pricing; our Grok 4.5 Cursor acquisition coverage tracks that, and the Grok 4.5 token-efficiency guide covers the tradeoffs.
Two things to watch. First, whether Meta's preview capacity holds under real load or introduces rate limits. Second, whether OpenAI leans on model routing and mixed-tier pricing — a pattern we examined in the GPT-5.6 routing guide.
How safe and reliable is Muse Spark 1.1?
Meta evaluated the model under its Advanced AI Scaling Framework v2, reporting improved resistance to jailbreaks, prompt injection, and hallucination. These are vendor-reported claims — validate on your own workload, especially for agent deployments where prompt injection through tool outputs is the dominant failure mode. Consumer privacy around Meta's AI ecosystem still applies; our Meta Muse and Instagram privacy guide covers the settings worth reviewing.
Should you switch to the Meta Model API today?
For most teams: run a benchmark, do not migrate blindly. The pricing is attractive and the compatibility endpoints make an A/B test straightforward. Muse Spark 1.1 fits high-volume agent pipelines, long-context document work, and cost-sensitive chat features. It is weaker for heavy code-modification where Opus 4.8 still leads, and for teams outside the US who cannot yet access the preview.
FAQ
Q: When did Meta launch its first paid AI API? A: 9 July 2026, alongside Muse Spark 1.1. It is the first time Meta has charged for API access to a frontier model.
Q: How much does the Meta Model API cost? A: $1.25/M input, $4.25/M output, $0.15/M cached input, and $2.50 per 1,000 web-search queries. New accounts get $20 in credits.
Q: Is Muse Spark 1.1 open-source like Llama? A: No. It is closed-weight and hosted-only. Llama remains Meta's open-weight family, but the frontier tier is now proprietary and paid.
Q: Can I use my existing OpenAI or Anthropic client code? A: Largely yes. The API supports both OpenAI-compatible and Anthropic-compatible request formats — most SDKs work with a base URL and key change.
Q: Where can I use Muse Spark 1.1 without paying? A: Free in "Thinking" mode in the Meta AI app and on meta.ai. The free tier does not extend to API calls.
Q: Is the Meta Model API available worldwide? A: Not yet. Public preview is US developers only at launch. Meta has not published a timeline for international rollout.
Discussion
0 comments