Meta has entered the commercial model API market with Muse Spark 1.1, a multimodal reasoning model that the company says is built for coding, computer use, and agent-style workflows. Just as important as the model itself, Meta is attaching aggressive usage pricing through a public preview of the Meta Model API, a move that could intensify competitive pressure on OpenAI, Anthropic, and xAI.

According to reporting from The Decoder and a wire item cited by Gadgets 360, Muse Spark 1.1 offers a 1 million-token context window and is available in “Thinking” mode in the Meta AI app and on meta.ai. The pricing reported by The Decoder is what stands out: $1.25 per million input tokens, $4.25 per million output tokens, and $0.15 for cached input, with web search grounding priced separately. If those terms hold broadly in production use, Meta would be setting a new low-price reference point among major U.S. API providers for near-frontier models.

For developers and enterprise buyers, this is not just another model launch. It is Meta moving from being primarily a supplier of open-weight models such as Llama into direct competition in hosted inference and developer platform economics. That matters because API pricing increasingly shapes which models get embedded into products, copilots, internal tools, and AI agents at scale.

What Meta is launching

The reported product package has two parts: Muse Spark 1.1 and the new Meta Model API. The model is described by The Decoder as a multimodal reasoning system aimed at agent-based tasks, programming, computer use, and multimodal understanding. Meta also reportedly positions the model as able to orchestrate multi-agent workflows, acting either as a coordinating “main agent” or as a focused subagent that hands off when needed.

The 1 million-token context window is another headline feature. In practical terms, that suggests the model is intended for long-running tasks that require large codebases, lengthy instructions, extensive tool traces, or multi-step workflow memory. Meta reportedly says the model can manage that large context by retrieving and compressing prior information while preserving important steps. That claim, if borne out, would make Muse Spark 1.1 more relevant for enterprise workflows where context persistence often matters as much as raw benchmark scores.

The launch also appears to mark a strategic shift in distribution. The Decoder notes that Muse Spark 1.1 does not ship with open weights, unlike Meta’s earlier Llama releases. If accurate, that indicates Meta is prioritizing a controlled API product over an open model release for this system. For builders, that changes the calculus: instead of self-hosting or customizing model weights, teams would evaluate Meta on hosted pricing, reliability, rate limits, and ecosystem fit.

Why the pricing is the real story

The biggest immediate impact comes from cost. The Decoder reports that Muse Spark 1.1 is priced below xAI’s Grok 4.5 and well below premium offerings such as OpenAI GPT-5.5 and Anthropic Opus 4.8 on output-token pricing. The article also mentions Fable 5 and Chinese alternatives such as GLM 5.2 as part of the broader pricing landscape.

That matters because output tokens are often where reasoning-heavy models become expensive in real deployments. A model that can think longer, use tools, and generate code or actions may be useful, but it also tends to consume more tokens. By lowering output-token pricing to $4.25 per million, Meta is signaling that it is willing to compete on unit economics, not only on benchmark quality.

The market implication is straightforward. OpenAI and Anthropic have built large businesses around premium API access to advanced models. Meta, by contrast, can afford to treat model APIs as part of a larger ecosystem strategy tied to Meta AI, meta.ai, and potentially broader platform distribution. The Decoder argues this could squeeze pure-play labs that depend more directly on model revenue and margins. That interpretation is plausible, though it remains market analysis rather than a confirmed operational effect.

Price cuts alone do not guarantee customer wins, but they do change procurement conversations. Enterprise AI teams evaluating inference spend now have another U.S.-based option that, at least on paper, reduces the cost of experimentation with long-context and agentic workloads.

The model claims and what is still unproven

Meta’s feature claims, as reported by The Decoder, are ambitious. The company says Muse Spark 1.1 can handle large enterprise codebases, diagnose complicated bugs, support feature additions, manage code migrations, and execute computer-use workflows across multiple applications. It also reportedly claims the model can decide when to click through interfaces directly and when to write scripts instead.

Those are attractive capabilities because they map closely to where AI spending is going: coding assistant products, internal automation, support tooling, and AI agents that can interact with software beyond chat interfaces. If Muse Spark 1.1 works reliably in those scenarios, it could be appealing to product teams building operational copilots or agent frameworks that need strong tool use and long context.

Still, the evidence presented in the source cluster is limited. The benchmark references in The Decoder include VALS-AI and Vibe Code Bench, where Muse Spark 1.1 is said to rank strongly or improve sharply over its predecessor. But those results are still being relayed through media coverage of Meta’s launch materials. Without independent reproducible evaluations, buyers should treat them as vendor-linked performance signals rather than settled proof.

The same caution applies to safety claims. The Decoder reports that Meta says it performed security evaluations under its Advanced AI Scaling Framework and that the model operates within acceptable ranges across frontier risk categories. That is useful context, but it is still Meta’s own assessment unless outside auditors or independent researchers publish corroborating findings.

Evidence, benchmarks, and competitive claims

This story rests primarily on reporting from The Decoder, with Gadgets 360 separately noting the key launch elements: Muse Spark 1.1, the 1 million-token context window, and the API preview. No primary Meta announcement text was included in the source material provided here, so the strongest product-performance and safety statements should be read as reported claims, not independently verified facts.

Several competitive comparisons in circulation also need careful framing. The Decoder says Muse Spark 1.1 undercuts Grok 4.5 and is far cheaper than OpenAI GPT-5.5 and Anthropic Opus 4.8 on output tokens. Those price comparisons are meaningful, but real-world cost depends on more than posted rates. Token efficiency, latency, tool-use overhead, caching behavior, and how much reasoning a model performs per task can materially change total spend.

The Decoder explicitly notes that lower listed prices may not translate into lower end-to-end costs if a model consumes more tokens or underperforms in production. That caveat is important. Enterprise buyers should compare task completion cost, not just token menu prices, especially for AI agents and coding assistant deployments where retries, long traces, and external tool calls can dominate expenses.

There is also a broader competitive claim embedded in the coverage: that Meta and Google can afford to use APIs strategically because they have larger businesses behind them, while independent labs face more pressure to maintain high margins. That is a reasonable market interpretation, but it is still interpretation. The immediate confirmed news is that Meta has launched the Meta Model API in preview and attached notably low pricing to Muse Spark 1.1.

What this means for builders and enterprise buyers

For builders, the launch creates a new option for long-context, reasoning-heavy applications without immediately forcing a move to Chinese open models or premium-priced closed APIs. Teams working on AI agents, browser and desktop automation, or repository-scale programming workflows may find the pricing attractive enough to run pilots they previously considered too expensive.

For enterprises, the decision will likely come down to four factors beyond headline cost. First is reliability: how consistently does Muse Spark 1.1 finish complex multi-step tasks? Second is integration: how well does the Meta Model API support existing orchestration stacks, MCP servers, and observability tools? Third is governance: what controls, logging, and data handling guarantees does Meta offer in preview? And fourth is model behavior under load: long context is only useful if latency and failure rates remain manageable.

The launch also raises strategic questions around Llama. If Meta is leaning harder into proprietary hosted systems for frontier capabilities, developers may need to separate Meta’s open-model story from its commercial API story. That would be a meaningful shift for a company that earned significant goodwill from open-weight releases.

What to watch next

The next signal to watch is whether Meta publishes fuller technical documentation and benchmark methodology for Muse Spark 1.1. Independent testing on coding, tool use, and computer-use tasks will matter more than launch-day comparisons.

Second, watch for enterprise adoption indicators around the Meta Model API. Named customers, integration partners, or support in common developer platforms would say more about market traction than posted token rates alone.

Third, pricing responses from OpenAI, Anthropic, and xAI will be important. If rivals cut prices, introduce lighter-weight tiers, or bundle more tooling around premium models, that would confirm Meta has changed competitive behavior rather than merely made a loud entrance.

Finally, watch whether Muse Image and other Meta models join the API. A broader commercial stack would make Meta more credible as a full platform supplier rather than a one-model disruptor.

Creati.ai perspective

Meta’s move looks less like a simple model release and more like a pricing attack on the hosted AI stack. Muse Spark 1.1 may or may not prove superior in production, but the Meta Model API changes the market by giving builders another major-provider option for long-context, agent-oriented workloads at a noticeably lower published price.

The bigger issue is strategic. When Meta, Google, and low-cost open-model ecosystems all push inference prices down at once, standalone labs lose room to charge premium rates unless they maintain a clear quality or workflow advantage. For product teams, that is good news in the short term: cheaper experimentation, more supplier leverage, and better odds that AI agents and coding assistant features can reach sustainable margins. The caveat is execution. If Muse Spark 1.1 cannot match its reported strengths on real software and automation tasks, low pricing alone will not win durable share.