Introducing Muse Code and Muse Spark 1.2
Simon Willison's Weblog Simon Willison ● Covered by 4 sources
Meta launched Muse Spark 1.2 and a new coding agent called Muse Code, tuned for long multi-step coding tasks. The wild part: it's way cheaper if you let Meta train on your data.
Meta just pushed out Muse Spark 1.2 alongside a companion tool called Muse Code, and the framing says a lot about where the model race has actually gone. It's not really about raw benchmark scores anymore. It's about whether a model can chain together long sequences of tool calls without falling apart, and Meta built an entire coding agent just to prove Spark 1.2 can do that.
Under the hood, Meta poured a lot more training compute into coding-specific tasks and widened the variety of environments the model learned from, while still holding onto its general-purpose agent skills from version 1.1. The two systems, Muse Code and Muse Spark 1.2, were trained together on purpose. Meta used rejection-sampled trajectories and tuned the recipe around goal-setting, context compaction, and subagent coordination, baking in the Muse Code toolset so the pairing works smoothly rather than as two bolted-together products. The training diet also leaned hard into long-horizon work: generating whole repositories, building large end-to-end projects, even automated research tasks that stretch far beyond a single prompt-response exchange.
Simon Willison, who tracks these releases obsessively, ran his usual pelican-riding-a-bicycle SVG test and called the 1.2 output a small but real step up from the July version. It's a silly benchmark on its face, but it's become a useful shorthand for whether a model's raw generation quality is actually moving.
The real story, though, might be the pricing. Meta split Spark 1.2 into two model IDs with wildly different rates. Standard muse-spark-1.2 costs $1.25 per million input tokens and $4.25 per million output tokens, putting it near Gemini 3.6 Flash's $1.50/$7.50. But there's a second version, muse-spark-1.2-contributor, priced at just $0.10/$0.20 — a discount so steep it undercuts even budget options like GPT-5.6 Luna and Gemini 3.1 Flash-Lite. The catch: you have to let Meta use your usage data to improve its products. Willison added both price points to llm-prices.com for easy comparison.
My take
Charging people ten times more for privacy is a business model dressed up as a discount, and it's worth calling that out plainly instead of treating it as a clever pricing tier. Meta isn't offering a deal, it's putting a price tag on your data and letting you choose whether you can afford to keep it. Expect every other lab to quietly test this same trick once they see nobody objects loudly enough.
Read more about this at: Simon Willison's Weblog
