Meta's Muse Spark 1.1 enters the paid AI coding market at $1.25 per million input tokens

Meta is putting a price tag on its AI ambitions. The company launched Muse Spark 1.1 on July 9, opening a public developer preview API for its most capable model yet at $1.25 per million input tokens and $4.25 per million output tokens, according to CNBC. New accounts start with $20 in free credits. The release marks Meta's clearest move yet into the paid, proprietary AI model market that Anthropic and OpenAI have built their businesses on.

A deliberate bet on coding and agentic workflows

Meta Superintelligence Labs, the unit led by AI chief Alexandr Wang, trained Muse Spark 1.1 specifically to excel at coding-related tasks because coding capability is foundational to building effective AI agents, Wang told CNBC. The model is designed to work with popular developer harnesses and third-party coding tools, a direct response to the surge in agentic AI adoption that accelerated across the industry in the first half of 2026.

The practical logic is straightforward: enterprise development teams increasingly rely on AI agents to handle multi-step software tasks autonomously. A model that integrates cleanly with existing tooling reduces the switching cost for those teams. Wang said the training approach was chosen to maximize adoption by meeting developers where they already work.

Muse Spark 1.1 is Meta's second notable AI release this week. Two days prior, the company released Muse Image, a model for generating images targeting creators and advertisers, according to CNBC. The pace signals a deliberate effort to fill out a model portfolio across multiple enterprise use cases simultaneously.

Pricing designed to scale with high consumption

The token pricing structure is built for volume users. At $1.25 per million input tokens, Meta is positioning Muse Spark 1.1 as an economically attractive option for organizations running high-throughput coding pipelines or large-scale agentic deployments. Wang described the goal as offering pricing that scales favorably as consumption grows, per CNBC.

For procurement and IT teams evaluating AI model vendors, the per-token structure is familiar, but the entry point matters. The $20 free credit offer lowers the evaluation barrier enough for engineering teams to run real workload tests before committing budget. Access is currently gated through a waitlist on Meta's developer portal, with early partners already live on the API.

A strategic turn from open source to revenue

Meta built its earlier AI reputation on the Llama family of open-source models, which attracted wide developer adoption but generated no direct revenue. Muse Spark represents the other side of that bet: a proprietary model sold as a service, running on Meta's own compute infrastructure. Wang told CNBC the model will be served on Meta's own systems, keeping the workload off third-party cloud platforms and marketplace aggregators like OpenRouter, at least for now.

Wang did indicate that an open-source variant of Muse Spark is in development, though he declined to give a release timeline to CNBC. A more powerful model, internally code-named Watermelon, is also in training. The current model carried the code name Avocado during development.

The commercial push reflects broader pressure on Meta to generate returns on its AI infrastructure spending, which rivals that of major cloud hyperscalers. Unlike those competitors, Meta does not currently operate a cloud business, though CNBC has reported the company is developing plans to enter that market. Muse Spark 1.1 and its API pricing are the most concrete evidence yet that Meta is building a model monetization engine alongside its infrastructure.

What this means for your team

• Evaluate API fit now: Muse Spark 1.1's waitlist is open. Engineering and AI procurement leads should apply early to test input/output costs against current Anthropic or OpenAI spend on coding workloads.
• Audit toolchain compatibility: Meta trained the model to work with popular developer harnesses. Confirm whether your team's existing agentic orchestration tools are on that list before committing to integration work.
• Model the token economics at scale: at $1.25 input and $4.25 output per million tokens, run your actual production volumes through the pricing before signing any enterprise agreement. High-output workloads, like code generation, will weigh heavily on the output rate.
• Watch the infrastructure constraint: API access is limited to Meta's own properties for now. Teams that depend on OpenRouter or multi-model aggregation layers will need to plan for a direct integration or wait for broader distribution.