Meta Launches Muse Spark 1.2 and Muse Code, Its First Coding Agent (August 2026)

On 5 August 2026 Meta's Superintelligence Labs released Muse Spark 1.2, an updated model aimed at coding and agentic workloads, together with Muse Code, the company's first coding agent, which runs on the new model and entered beta the same day. Meta benchmarked Spark 1.2 against Anthropic's Opus 5, OpenAI's GPT-5.6 Terra, Google's Gemini 3.6 Flash and xAI's Grok 4.5, according to Yahoo Finance's coverage of the launch.

Model trackers reported a 1 million token context window with context compaction and asynchronous tool calling, and API pricing of $1.25 per million input tokens on the standard tier, with a heavily discounted "contributor" tier at $0.10 per million for customers who allow their usage to feed training. Meta says the model was co-trained with the Muse Code multi-agent harness, meaning the agent and the model were tuned together rather than the agent being layered on afterwards. Five days later, on 10 August, CNBC reported Meta's release of Muse Glimmer, an open-weight model, signalling that the company has not abandoned open distribution despite the shift away from the Llama brand.

The financial backdrop is significant. Meta raised its infrastructure spending guidance to between $134 billion and $145 billion for the year after Llama's underwhelming reception, hired Scale AI's chief executive as chief AI officer, and is developing Muse Image and Muse Video in parallel. Its stock had fallen more than 20% over the preceding year.

Why it matters

Coding agents are now the primary competitive battleground among frontier labs, and Meta's entry follows OpenAI's ChatGPT Work and Anthropic's Claude Code by months, not years. For enterprises, the more interesting design choice is the contributor tier: a twelvefold price cut in exchange for data rights. That is a procurement decision with data-governance consequences, not merely a pricing option.

The co-training of model and agent harness also raises a portability question. If Spark 1.2 performs best inside Muse Code, benchmarks run through other agent frameworks may understate or overstate its value for your stack. Elsewhere in the open-weight market, DeepSeek released its V4-Flash model officially on 31 July after a preview that had topped OpenRouter usage for seven weeks, with API prices cut by up to half, so price competition is arriving from several directions at once.

What it means for leaders

• Read the contributor-tier terms before anyone selects it. Discounted tiers that license your prompts and code for training can conflict with confidentiality clauses, GDPR purpose limitation and ISO/IEC 42001 data-governance controls.
• Govern coding agents as privileged users. Muse Code and its peers commit code and run tools; apply the same review gates, secret scanning and branch protections you would to a contractor.
• Evaluate in your own harness. Ask vendors for results outside their proprietary agent framework, and run internal evaluations that reflect your repositories and languages.
• Watch the open-weight track. Muse Glimmer and DeepSeek V4-Flash change the cost floor; open weights also require your own safety testing and licence review before deployment.
• Factor in vendor volatility. Meta's spending and strategy shifts are rapid; keep model abstractions in your architecture so that switching is a configuration change.