Meta has released Muse Spark 1.1, an upgraded multimodal reasoning model built for coding and agentic tasks, alongside the company’s first-ever paid developer API. The launch marks a significant strategic pivot for a company that built its AI reputation on open-weight models, and signals Meta’s intent to compete directly with Anthropic and OpenAI in the fast-growing market for AI coding and automation tools.
From Open Weights to a Paid API
Until now, developers accessed Meta’s AI models primarily through open-weight releases under the Llama family. Muse Spark 1.1 breaks that pattern, launching on the new Meta Model API in public preview, a self-serve platform that charges developers directly for access, the same commercial model long used by Anthropic and OpenAI. The API is priced at $1.25 per million input tokens and $4.25 per million output tokens, positioning it above OpenAI’s entry-level GPT-5 mini and Anthropic’s Claude Haiku 4.5, but below Anthropic’s higher-end Claude Sonnet 4.6.
The public preview is currently limited to developers in the United States, meaning Indian developers and enterprises will need to wait for wider regional rollout before they can build directly on the model, a detail worth noting given India’s position as one of the largest developer markets globally and a key growth region for AI coding tools such as GitHub Copilot and Claude Code.
A Sharp Jump in Coding and Agentic Capability
Muse Spark 1.1 succeeds the original Muse Spark, which Meta’s Superintelligence Labs released in April as its first text and reasoning model. The upgrade specifically targets coding, an area where the earlier version lagged rivals; independent benchmarking firm Artificial Analysis had scored Muse Spark 1.0 at 59.0 on coding tasks compared with 80.8 for Anthropic’s Claude and 75.1 for OpenAI’s GPT-5.4.
Meta says the new version can diagnose and fix complex bugs, implement features in large enterprise codebases, and execute large-scale code migrations, while supporting agentic coding features such as planning mode, subagent delegation and context compaction. The model also carries a one-million-token context window and can coordinate multiple subagents working in parallel to reduce task latency, alongside expanded multimodal capabilities that let it interpret images, video, audio and PDFs while operating tools or software on a user’s behalf.
Meta’s Chief AI Officer, Alexandr Wang, said the model performs competitively against GPT-5.5 and Anthropic’s Opus 4.8 on several agentic evaluations. Independent benchmarking site Vals AI separately reported that Muse Spark 1.1 topped its MedScribe and TaxEval leaderboards and secured the top position on Harvey’s Legal Agent Bench, while being considerably cheaper and faster than competing frontier models, though such vendor and third-party benchmark claims typically warrant independent verification before enterprise adoption decisions.
Early Partners and Safety Disclosures
Meta named Replit, Cline and Box as early API partners. Replit’s chief executive highlighted the model’s context window and multimodal support, while Cline’s chief executive pointed to its tool-use capabilities and pricing for large-scale coding workloads. Box’s vice-president of AI products said the model performed competitively against frontier alternatives on the company’s internal enterprise evaluation set.
On safety, Meta said Muse Spark 1.1 was assessed under its Advanced AI Scaling Framework, which evaluates frontier risks including chemical and biological misuse, cybersecurity threats, and loss of control. The company’s evaluation report noted that the unmitigated model reached a high-risk threshold in the chemical, biological and cybersecurity categories, but that layered safeguards brought residual risk down to what Meta assessed as moderate or lower before release. The company also reported improved resistance to jailbreak attempts and prompt-injection attacks compared with the earlier version.
The model is also replacing existing Llama models powering Meta’s consumer AI assistants across WhatsApp, Instagram, Facebook and its smart glasses line, extending its reach well beyond the developer community targeted by the new API.
