Mark Zuckerberg broke a years-long silence on X this week. He announced Muse Spark 1.1. The model isn’t just another upgrade. It marks Meta’s first serious charge for developer access to a flagship AI system.
Hours after the July 9 launch, analysts and developers scrambled to test it. Results poured in. The model excels at agentic tasks. Those are the multistep operations where software acts with minimal human guidance. It writes code, debugs systems, orchestrates tools across apps, and processes video, images, and documents in one pass. Android Authority first highlighted how Meta aims to undercut ChatGPT and Gemini on cost while delivering competitive performance.
Yet price tells only half the story. Muse Spark 1.1 carries a one-million-token context window. That’s enough memory to hold entire codebases or long-running enterprise workflows. The model actively compacts stale information. It keeps what matters for continuity. Such efficiency matters when agents run for hours or days.
Meta built this version inside its Superintelligence Labs. The unit formed last year under chief AI officer Alexandr Wang. The original Muse Spark debuted in April. It focused on multimodal reasoning and basic tool use. Version 1.1 sharpens every edge. Gains appear in coding accuracy, computer control, and resistance to hallucinations. Independent tests from Artificial Analysis show it scoring 51 on their Intelligence Index, up eight points in three months. It sits near models from OpenAI and Anthropic but at far lower cost.
One dollar and twenty-five cents. That’s the input price per million tokens. Output runs four dollars and twenty-five cents. Cached inputs drop to fifteen cents. Those numbers land roughly one-quarter the rate of comparable frontier models. Zuckerberg called the approach “aggressive” in a Bloomberg interview. He intends to make high-capability AI available to more businesses. Not just the ones with deep pockets.
The shift carries financial weight for Meta. The company has poured tens of billions into compute infrastructure. Investors questioned the return. Free consumer tools on Instagram, WhatsApp, and the Meta AI app remain. The new Meta Model API opens a paid tier for developers. It creates a fresh revenue stream without alienating everyday users. Bloomberg reported the move as Meta’s entry into charging for AI services. Read the full story here.
But does cheap win the enterprise? TechCrunch examined the crowded coding arena. Muse Spark 1.1 targets multistep reasoning across external services. It fixes bugs. It migrates large codebases. It manages digital workflows. Enterprises crave exactly that automation. “Muse Spark 1.1 delivers exceptional performance in personal agentic tasks that require planning and orchestration across a range of external apps and services,” Meta stated in its announcement blog. The company positioned the model to replace Llama instances inside its own apps and glasses.
Real-world tests back some claims. On professional tool-use benchmarks like JobBench and MCP Atlas, it leads. Pure coding benchmarks such as SWE-Bench still see Claude Opus 4.8 and certain GPT variants ahead. The gap narrows fast. Rowan Cheung noted on X that many doubted Meta’s position in the AI race. “Yesterday, they dropped Muse Spark 1.1, now one of the strongest agentic models, and massively undercut OpenAI and Anthropic on price.” He recalled interviewing Zuckerberg last year. The CEO stressed building the highest compute per researcher and doing whatever it takes to expand capacity. Those bets now show returns. Meta shares rose more than 10 percent following the release.
The timing feels deliberate. OpenAI, Anthropic, and Google have spent years cultivating developer loyalty through sophisticated APIs and ecosystems. Meta arrives late yet arrives armed with scale. Its social platforms feed enormous training data. Its hardware investments yield dense clusters. And its willingness to price low disrupts assumptions about what frontier intelligence must cost.
Reuters covered the debut in detail. The wire service emphasized how the upgraded model writes and debugs code, understands multiple media types, and completes complex tasks with less intervention. It also powers a “Thinking” mode in the Meta AI app. That mode lets the system reason step by step before answering. Early users report tighter focus and fewer detours. Reuters story from July 9.
New York Times reporters framed the launch inside the global technology race. They noted Meta’s departure from its longtime practice of giving AI away for free. The paid tier targets businesses ready to embed agents into operations. Fortune added context on benchmarks. Muse Spark 1.1 beats Google’s latest Gemini in coding and reasoning according to Meta’s tests. It surpasses older versions of rival models in select categories. The publication tied the release to Wang’s leadership of the Superintelligence Labs. Three months separate the first Muse Spark from this iteration. Progress arrives quickly when resources align.
Developers on X traded early impressions Thursday night. Some praised the model’s ability to maintain context across million-token sessions. Others tested it on finance workflows and software engineering problems. Token efficiency stood out. Artificial Analysis calculated roughly twenty-six cents per full Intelligence Index task at Meta’s pricing. That figure sits well below several competitors. Speed hits about 114 output tokens per second on the company’s API. Time to first token averages twenty-one seconds. Respectable for a system that thinks before it speaks.
Risks remain. The model still trails on certain knowledge-work evaluations. Hallucination rates dropped but have not vanished. Meta evaluates against chemical, biological, cybersecurity, and loss-of-control categories under its own scaling framework. The company publishes some results. Full transparency on training data and methods stays limited compared with its earlier open-weight efforts.
Analysts wonder whether price pressure will force OpenAI and Anthropic to respond. Recent weeks already brought cheaper tiers from those firms. Yet none match Muse Spark 1.1’s combination of agentic strength and sub-five-dollar output pricing. DataCamp’s breakdown called it Meta’s second Superintelligence Labs model. The publication stressed the 1-million-token window and native multimodal design. It also flagged built-in web search grounding. Add one tool call and the model returns cited, real-time answers without extra retrieval plumbing.
Inside Meta the model will spread. It replaces older Llama versions across WhatsApp, Instagram, Facebook, Messenger, and the Ray-Ban smart glasses. Consumer chats gain sharper reasoning. Voice responses quicken. Shopping and support features grow more capable. The same underlying system that enterprises pay to access will improve daily experiences for billions of users. That dual track separates Meta from pure AI startups.
Zuckerberg once joked that his company would keep building bigger models until someone told him to stop. No one has. The Superintelligence Labs roadmap points toward personal superintelligence. Each Muse generation validates the last before the next leap. Muse Spark 1.1 feels like proof of concept and product at once. It demonstrates that Meta can compete on capability. It proves the firm will compete on cost.
Watch the next months closely. Enterprise pilots will reveal where the model shines and where it still needs work. Competitors will adjust pricing or accelerate releases. Developers will route traffic among multiple APIs based on task and budget. The era of measuring AI solely by benchmark scores fades. Cost per completed task now decides winners.
Meta just lowered the bar. Others must clear it or step aside. The AI race didn’t slow. It simply changed lanes.
WebProNews is an iEntry Publication