Breaking

Meta rolls out Muse Spark 1.3, leading GPT-5.6 Sol and Claude Opus 5 on DeepSWE v1.1

September 2, 2026 at 15:35 EDT

Foundation Models
Software Dev & Coding
AI Agents

Meta Superintelligence Labs released Muse Spark 1.3 on September 2, 2026, rolling the coding- and agent-focused model out in Muse Code and the Meta Model API, and publishing a scorecard that places it above GPT-5.6 Sol and Claude Opus 5 on DeepSWE v1.1.

September 2, 2026 · Meta

Muse Spark 1.3 hits the coding frontier

Meta Superintelligence Labs is rolling out Muse Spark 1.3 in Muse Code and Meta Model API. Its launch scorecard places the model ahead of GPT-5.6 Sol and Claude Opus 5 on DeepSWE v1.1.

75.4

DeepSWE v1.1 (max)

88.8

Terminal-Bench 2.1, tied with GPT-5.6 Sol

98.1

MRCR v2, 512K–1M band

$1.25

Standard input per 1M tokens ($4.25 output)

DeepSWE v1.1, drawn to scale

Task pass rate on Meta's launch chart. Column height is proportional to each score.

55.0

Spark 1.2
xhigh

73.0

GPT-5.6 Sol
max

74.0

Opus 5
max

75.4

Spark 1.3
max

A cheaper loop, not a longer grind

Meta engineers report about 20% fewer tool calls versus Muse Spark 1.2. Five blocks versus four is that cut, drawn as stacked load.

1.2 tool load
5 units

1.3 tool load
~20% fewer

Internal comparisons also found about 25% fewer tokens, with fewer unnecessary turns and less verbose diffs.

What improved

Leads DeepSWE v1.1 and SWE-Atlas Codebase QnA on Meta's chart, ties the Terminal-Bench 2.1 top score at 88.8, and is described as a tighter coding loop at the same $1.25 / $4.25 standard rate.

What still lags

Opus 5 still leads GDPVal-AA v2, OSWorld 2.0, JobBench, and AutomationBench. Scores are vendor-reported; max reasoning is not live yet; audio understanding in 1.3 may be degraded.

Live today, with two caveats

Available

Muse Code and Meta Model API, model ID muse-spark-1.3

Not yet

Max reasoning waits on additional safety testing

Watermelon and Muse Spark open weights, no date given

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.