Meta launched Muse Spark 1.1 on July 9, marking the company's first paid AI model after years of free Llama releases, representing a sharp shift from Meta's previous open-source strategy. Meta priced the new model at $1.25 per million input tokens and $4.25 per million output tokens, with AI chief Alexandr Wang characterizing the pricing as "very aggressive and attractive" compared to competitors. Wang explained that the goal is to have "attractive pricing that scales with immense consumption usage," positioning the company against OpenAI, Anthropic, and Google in the competitive AI market. The company charges $1.25 per million input tokens and $4.25 per million output tokens, making Muse Spark 1.1 cheaper than SpaceXAI's new Grok 4.5 and undercuts Anthropic's Opus model. New accounts receive $20 in free credits before billing starts, with access limited to US developers requiring a waitlist and staying off third-party marketplaces like OpenRouter. The pricing gap could matter most for heavy, high-volume workloads, with a developer running the same task across all four models paying Meta's rate for less than a third of what GPT-5.5 charges on output alone.
Muse Spark 1.1 represents Meta's "strongest model for agentic and coding work yet," as reported by CNBC, with Wang emphasizing that the model was trained to excel in coding-related tasks to improve overall agentic capabilities. The model can plan tasks, use software and tools, and operate a computer across desktop, mobile, and browser platforms, handling long, multi-step work and holding up to one million tokens in memory. One of the biggest upgrades is its ability to work as an AI agent that can plan, organise, and complete tasks across different apps and services, according to Meta's announcement. The model can create a plan, collect the required information, and even divide work among multiple AI sub-agents to complete tasks more efficiently. Meta says Muse Spark 1.1 performs better in computer-use workflows that involve several applications and changing information, deciding when it is faster to automate a task using scripts and when direct interaction with the interface is more practical.
Muse Spark 1.1 demonstrates strong performance in tool-use benchmarks, particularly excelling in agentic tasks, according to Meta's latest specifications. The model achieved an MCP Atlas scaled tool use score of 88.1, ahead of Opus 4.8 and GPT-5.5, while JobBench professional tool use reached 54.7 versus Opus 4.8's 48.4 and GPT-5.5's 38.3. However, Terminal-Bench 2.1 coding performance scored 80.0, trailing GPT-5.5's 83.4 and Opus 4.8's 82.7. The model's OSWorld-Verified benchmark reached 80.8, again behind Opus 4.8's 83.4. Meta's pricing strategy includes a $0.15 cached input rate, which is notably lower than most competitors' standard input rates, with generous rate limits including free tier access to 60 requests per minute with 2 million tokens per minute, and paid tier offering 3,000 requests per minute with 4 million tokens per minute. Meta's pricing is higher than OpenAI's GPT-5 mini and Anthropic's Claude Haiku 4.5, but cheaper than Anthropic's Claude Sonnet 4.6, according to Reuters. The gap narrows once Sonnet 5's introductory pricing expires on August 31, though Meta still comes out cheaper on both ends even against the standard $3/$15 rate.
Meta has officially moved from AI evangelist to AI toll collector, launching its first serious paid developer API after years of free Llama releases, as reported by Investing.com India. According to Zuckerberg, Muse Spark 1.1 will be "among the most affordable options" in the market, with pricing reportedly around 25% of the cost of comparable frontier models from OpenAI and Anthropic. The real weapon is price - if Muse Spark 1.1 is offered at roughly 25% of rival frontier models, the AI race just became a margin war, according to Investing.com India. This is no longer just a model race - it is becoming an AI price war, with Meta walking in with a discount hammer and daring the incumbents to defend margin in a market where everyone is already spending like the future has been mortgaged twice. The problem for AI bulls is that growth and pricing power are not the same thing - if Meta can offer near-frontier capability at a quarter of the price, then OpenAI, Anthropic, Google, xAI, and Chinese labs all have to answer the same ugly question: how much intelligence can you sell before intelligence starts trading like bandwidth.
Several companies have already started using the API during its early rollout, with the first major partners including Replit, Cline, and Box. Replit CEO Amjad Masad said the combination of a one-million-token context window, multimodal abilities, and advanced coding features could change how developers build software at scale, according to Investing.com. Cline CEO Saoud Rizwan said the model's flexible tool use and competitive pricing make it attractive for developers handling large coding workloads. Box Vice President of AI Products Yashodha Bhavnani said Muse Spark 1.1 performed on the same level as some of the industry's leading AI models during Box's internal testing, according to Investing.com. Meta plans to replace the current Llama AI models used in WhatsApp, Instagram, Facebook, and Meta smart glasses with Muse Spark, indicating the model's strategic importance across Meta's platform ecosystem. Version 1.1 is available now in "Thinking" mode through the Meta AI app and website, with the model already powering a wide range of features across Facebook, Instagram, WhatsApp, and the Meta AI app.
Meta carried out extensive safety testing before releasing Muse Spark 1.1, looking for risks involving chemical and biological misuse, cybersecurity threats, and situations where AI systems could lose control. Meta said Muse Spark 1.1 stayed within its safety limits and that a detailed safety report will be released soon. While Meta previously emphasized releasing its Llama family models to the open-source community, the company is now focusing on selling access to proprietary AI models as part of its strategic pivot under Alexandr Wang's leadership. Wang confirmed Meta remains "committed to open source" and that MSL has a "variant of Muse Spark that is in development that we do intend to open source," though he declined to specify timing. Meta is currently training a more powerful AI model, code-named Watermelon, but didn't say when it would be released, indicating continued investment in AI development despite Wall Street pressure for returns on massive AI infrastructure investments. Longer term, Wang is pursuing a vision of a more agentic Meta AI that could handle tasks like planning a party or a vacation and take more direct action on its own, seeing Meta's strength in its billions of users and how much the company knows about them: "That's really the part that no one else can replicate."
