Muse Spark 1.2

Muse Spark 1.2 is positioned by Meta as the model behind Muse Code, a beta terminal-based coding agent for macOS and Linux that takes on complex software engineering work across large repositories. According to Meta's research blog, the agent plans changes, writes code, and validates results, coordinating multiple persistent async background subagents that remain active throughout each session rather than being spawned per task. This persistent-subagent design is intended to reduce redundant information gathering and lower the steering burden on difficult, multi-step jobs.

The runtime built around Muse Spark 1.2 uses a local event log that records every model call, tool run, approval, and edit, creating a single replay-exact and restart-safe source of truth. After a crash, the agent can resume precisely where it stopped, which lets it sustain long-running tasks without being derailed by failures. Together, those design choices frame the model as a practical fit for developers who want an autonomous coding assistant capable of handling sustained repository-scale engineering work with minimal interruption.

Abacusmuse-spark-1.2muse

Open provider page Provider docs

Quick Info

Provider: Abacus
Model key: muse-spark-1.2
Release date: Aug 5, 2026
Last updated: Aug 5, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $1.25
Output token cost: $4.25

Limits

Output tokens: 131,072 tokens
Context window: 1,048,576 tokens

Transparent token rates

Compare Muse Spark 1.2 pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Muse Spark 1.2

Abacus

Coverage

Digital Applied's launch guide covers the Muse Spark 1.2 release on August 5, 2026, reporting that Meta shipped the model alongside Muse Code in beta. The piece details a 1,048,576-token (1M) context window and two published Meta Model API price tiers: $1.25 input / $4.25 output per million tokens on the standard tier, The guide notes that benchmark charts were published as images rather than tabular data, limiting independently verifiable score detail, and that launch coverage focused disproportionately on pricing. The model is attributed to Meta (via Meta Superintelligence Labs), not to Abacus; Abacus's RouteLLM gateway role as a s

Abacus

CoverageBenchmark

Artificial Analysis published an independent benchmark of Muse Spark 1.2 (xhigh) on August 5, 2026, awarding it an Intelligence Index score of 54, up three points from Muse Spark 1.1's 51 and eleven points from Muse Spark 1.0's 43. The piece places 1.2 effectively tied with GPT-5.5 (xhigh, 55) and Grok 4.5 (high, 54), Agentic performance improved materially: GDPval-AA v2 Elo rose 260 points to 1631 (5th among all models benchmarked, ahead of Claude Opus 4.8 max at 1588), Terminal-Bench v2.1 gained two points to 80%, and τ³-Banking rose two points to 27%. Cost-per-task landed at $0.40 on the unchanged $1.25/$4.25 pricing, making 1.2

Abacus

CoverageBenchmark

BenchLM's aggregator profile directly addresses Muse Spark 1.2 as a named version, reporting a capability composite of 70.4/100 against a field median of 56.2. It lists published API pricing at $1.25 input / $4.25 output per million tokens, a cached input rate of $0.15 and a blended rate of $2.75. The page records thro Category percentile standings place Muse Spark 1.2 at the 89th percentile for Agentic (17 of 153 ranked), 87th for Coding (21 of 152), and 95th for Knowledge (10 of 183). The page flags Muse Spark 1.2 as superseded by Muse Spark 1.3 and does not address the Abacus RouteLLM gateway specifically; the model is presented a