What is Muse Spark 1.2?

Muse Spark 1.2 is Meta Superintelligence Labs' coding-focused large language model, released August 5, 2026, with a 1-million-token context window. It's Meta's third Muse Spark model in four months and was co-trained alongside Muse Code, a terminal coding agent. API pricing is $1.25 input / $4.25 output per million tokens ($0.15 cached), or $0.10/$0.20 on the Contributor tier that trains on your data.

Muse Code is Meta's first dedicated terminal coding agent, launched in beta on August 5, 2026, powered by Muse Spark 1.2. It runs on macOS and Linux, coordinates persistent subagents, and uses a replay-exact, restart-safe event log so crashed tasks resume exactly where they stopped.

How much does Muse Spark 1.2 cost?

Standard API pricing is $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15. Meta also offers a Contributor tier at $0.10/$0.20 per million tokens — the cheapest of any frontier coding model — in exchange for using your prompts and completions to train Meta products.

Quick Answer

What Is Meta Muse Spark 1.2 & Muse Code? (August 2026)

Published: August 9, 2026

The Short Answer

Muse Spark 1.2 is Meta Superintelligence Labs’ coding-focused LLM, released August 5, 2026 with a 1M-token context window. It shipped alongside Muse Code — Meta’s first dedicated terminal coding agent (beta, macOS/Linux). Standard API pricing is $1.25/$4.25 per MTok, with a headline-grabbing Contributor tier at $0.10/$0.20 that trains on your data.

Muse Spark 1.2: The Model

This is Meta’s third Muse Spark model in four months (after the original Muse Spark and Muse Spark 1.1 on July 9). The 1.2 release is coding-focused and was co-trained with Muse Code so the model and agent work as a pair. It has a 1,048,576-token context window and is positioned as a step toward Meta’s frontier ambitions, with larger models promised.

Muse Code: The Agent

Muse Code is a terminal-only agent (no desktop app) built for large-repo software engineering. Its distinguishing features:

Replay-exact, restart-safe event log — an append-only log means a crashed session resumes at the exact step it stopped.
Persistent subagents — multiple agents stay active across a session to plan, write, and validate.
Bundled skills — /plan (approval-gated plans), /grill (stress-test a plan), /goal (work toward an objective).

Pricing: The Two-Tier Play

Tier	Input	Output	Catch
Standard	$1.25 / MTok	$4.25 / MTok	Cached input $0.15; not used for training
Contributor	$0.10 / MTok	$0.20 / MTok	Meta trains on your prompts + completions

The Contributor tier is the most aggressive pricing from any frontier coding model — cheaper than DeepSeek V4 Flash — but only appropriate for non-sensitive code you’re comfortable donating to Meta’s training set.

Benchmarks & Reality

Independent Vals testing ranked Muse Spark 1.2 mid-pack on TerminalBench but strong on value per test (~$0.69). It’s a capable, cheap coding model, not a benchmark leader against Claude Opus 5 or GPT-5.6 Sol.

Why It Matters

Meta’s post-Llama pivot to the Muse Spark family now has its own coding agent, putting it directly against Claude Code, Codex, and Cursor. The Contributor pricing is a classic Meta move: subsidize adoption with data. For cost-sensitive teams doing non-proprietary work, it’s the cheapest agentic coding on the market.

Sources

MarkTechPost — Meta releases Muse Code (Aug 5, 2026): marktechpost.com
DataNorth — Meta releases Muse Spark 1.2 and Muse Code: datanorth.ai
OpenRouter — Muse Spark 1.2 pricing & benchmarks: openrouter.ai