Muse Spark 1.3 Benchmarks

Meta published official Muse Spark 1.3 benchmarks on 2026-09-02, highlighted by a 75.4% solve rate on DeepSWE 1.1 and an 88.8% score on Terminal-Bench 2.1.

At Layer3Labs, we evaluate agentic coding models across real engineering pipelines, where tool calling speed and token burn matter as much as synthetic test accuracy.

The published data covers two release variants, Muse Spark 1.3 (max) and Muse Spark 1.3 (xhigh), alongside direct efficiency comparisons against Muse Spark 1.2.

Muse Spark 1.3 Benchmarks across Core Coding Suites

Meta published four core coding and retrieval benchmark scores for Muse Spark 1.3 to demonstrate its multi-step agentic capabilities on 2026-09-02.

Muse Spark 1.3 scored 75.4% on DeepSWE 1.1, which measures an agent's ability to resolve end-to-end software engineering issues across complete repositories.

On Terminal-Bench 2.1, the model achieved an 88.8% success rate, testing bash execution accuracy, command navigation, and environment tool control. Software teams can review detailed run definitions directly on the Meta AI Research announcement.

For codebase comprehension, the model scored 59.4% on SWEAtlas CodeBase QnA, which evaluates how accurately a model answers architectural and dependency questions across unfamiliar codebases.

Long-context retrieval reached 98.5% across its 1M-token context window, confirming high needle-in-a-haystack recall when parsing massive log files or multi-file diffs.

• DeepSWE 1.1: 75.4% (end-to-end agentic software issue resolution)
• Terminal-Bench 2.1: 88.8% (terminal environment navigation and bash command execution)
• SWEAtlas CodeBase QnA: 59.4% (repository-wide comprehension and architectural queries)
• Long-context retrieval: 98.5% (data recall across 1M-token context window)

Efficiency Metrics and Operational Token Reductions

Meta reported that Muse Spark 1.3 completes coding workflows using roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2.

Engineers at Meta Superintelligence Labs (MSL) tracked these efficiency gains on identical multi-step software tasks in Muse Code, Meta's command-line interface (CLI) agent.

Fewer tool calls mean the model generates more precise sub-actions, reducing iterative looping during debugging passes. The 25% reduction in overall token burn directly lowers execution latency during long continuous runs.

These efficiency gains occur while preserving the full 1M-token context window established in the previous generation. Teams upgrading from Muse Spark 1.2 will see shorter execution times on complex pull requests without losing repository memory.

• Tool call reduction: ~20% fewer external calls than Muse Spark 1.2
• Token volume reduction: ~25% fewer generated tokens on identical coding tasks
• Context window: 1M tokens preserved from the prior model generation
• Environment target: optimized for Muse Code CLI agent workflows

Intelligence Index Discrepancy across Third-Party Reports

Published evaluation records show differing numbers for the overall Artificial Analysis Intelligence Index score of Muse Spark 1.3.

Early financial press reporting from Bloomberg stated that the top-tier max variant earned a score of 62 on the index, positioning it behind only Claude Fable 5.1 and Claude Opus 5.

The per-variant release data hosted on the Artificial Analysis model page documents lower individual benchmark scores across specific capability categories.

These discrepancies stem from testing variations between partner preview builds and early public evaluations. Because published figures differ across sources, teams should check the live index tables directly at Artificial Analysis rather than relying on a single reported score.

• Reported index figure: 62 on the Artificial Analysis Intelligence Index according to Bloomberg reports
• Comparative standing: placed behind only Claude Fable 5.1 and Claude Opus 5 in early partner reporting
• Public tracker variance: per-variant breakdown pages display lower individual category scores
• Verification: verify current updated figures at Artificial Analysis

Release Variants and Closed-Weight Distribution

Meta structured the launch of Muse Spark 1.3 around two distinct performance tiers with different access paths.

The flagship tier is Muse Spark 1.3 (max), which delivers maximum reasoning capacity but remains in limited preview for selected Meta partners pending further safety testing.

The general production tier is Muse Spark 1.3 (xhigh), which powers standard deployments in the Meta Model Application Programming Interface (API) and the Muse Code agent.

Model weights remain strictly closed for both tiers. Rollouts will expand from developer tools into consumer surfaces across Meta AI, Instagram, and Facebook over subsequent release cycles.

Meta did not publish an itemized input and output pricing card on launch day. Independent trackers at Artificial Analysis noted an estimated blended rate of roughly $0.80 per million tokens. For context, Muse Spark 1.2 pricing split usage into a Standard tier at $1.25 input and $4.25 output alongside a Contributor tier at $0.10 input and $0.20 output, but Meta has not confirmed whether 1.3 will adopt that structure.

• Muse Spark 1.3 (max): limited preview for Meta enterprise partners, reasoning mode gated for safety testing
• Muse Spark 1.3 (xhigh): broadly accessible production tier for API and Muse Code integration
• Weights: closed model distribution with no local weights released
• Pricing status: unconfirmed rate card with an estimated blended rate of roughly $0.80 per million tokens

Interpreting Muse Spark 1.3 Benchmarks for Production

Synthetic scores validate high-end model reasoning, but development teams must measure performance inside their own software stacks.

When we evaluate coding agents for client infrastructure, we measure whether synthetic gains on benchmarks like Terminal-Bench translate into fewer broken builds in continuous integration.

This release is not suitable for organizations that require self-hosted open weights or immediate unvetted access to the highest reasoning tier. Teams needing fully private infrastructure should continue running local open-weight alternatives or Meta's sibling models like Muse Glimmer.

Our assessment would shift if Meta releases the max variant for general public API access and provides an unbundled per-token rate card with clear data-use terms. Until independent third-party audits reconcile the variant scores, teams should treat top-line marketing figures as upper bounds.

Review your own repository test results alongside the published benchmarks before migrating production agent pipelines.

• Synthetic vs real world: 75.4% DeepSWE demonstrates agent competence, but repository conventions dictate actual success
• Who this is not for: teams requiring local open weights, on-premise hosting, or ungated access to the max tier
• What would change our answer: public availability of the max variant and confirmed enterprise data privacy terms
• Recommended action: run dual-branch A/B tests on existing pull requests to measure token burn

How to use Muse Spark 1.3

A hosted model runs on the provider's servers, so using it is really about the tool you access it through.

The fastest way to put Muse Spark 1.3 to work day to day is inside an AI IDE, and Cursor is the most popular — it supports every major model, so you can be working in minutes. Each major maker also ships a first-party tool — Claude Code, Codex, or Antigravity — worth trying for the native experience. Prefer a different editor? Windsurf, Zed, and GitHub Copilot drive these models too.