Model record · checked 06 Aug 2026

Muse Spark 1.2 is the model, not the agent.

Meta describes Muse Spark 1.2 as a coding-focused update optimized for multi-file refactors, complex debugging, whole-repository context, and long-running agent workflows.

Independent guide · official facts separated from interpretation

01 · Identity

The inference layer under Muse Code.

Muse Code supplies the terminal runtime, tools, state, approvals, agents, and event log. Muse Spark 1.2 supplies model outputs. Meta says the model was co-trained with the harness but also generalized across other coding agents.

Standard model ID
muse-spark-1.2
Contributor model ID
muse-spark-1.2-contributor
Context window
1M tokens Meta says
Access
Muse Code, Meta Model API, and OpenRouter as listed by Meta
Release date
05 Aug 2026

02 · Pricing snapshot

Meta Model API list pricing as checked on 06 Aug 2026. Pricing can change.
TierCached inputInputOutputData / access qualification
Standard$0.15 / 1M$1.25 / 1M$4.25 / 1MMeta says prompts and completions are not used to train its models; zero-data-retention requests are beginning through sales.
ContributorSee current Meta termsSee current Meta termsSee current Meta termsToken-rate-limited in a rolling five-hour window, select countries, and data may be used to improve Meta products.

Reasoning tokens are billed as output, according to Meta’s launch guide. This site does not estimate cost per task without a reproducible workload.

03 · Training claims

Co-trained for the harness around it.

Meta says training included rejection-sampled harness trajectories and recipe optimization for goals, compaction, and subagents. Long-horizon training covered whole-repository generation, large end-to-end projects, and auto-research.

Claim classificationVendor-reported
co-training     harness trajectories
conditioning goals + planning
memory context compaction
workloads whole repositories

Qualification:
These are Meta’s training descriptions.
No weights, training corpus, or independent
reproduction is supplied by this site.

04 · Evaluation limits

Compare configurations, not logos.

A coding score belongs to a model, harness, effort setting, environment, date, and budget. Meta’s charts are useful release evidence; they do not substitute for an independent workload on your repository.