Breaking

Meta rolls out Muse Spark 1.3 with stronger coding and agentic results using about 25% fewer tokens

September 2, 2026 at 20:31 EDT

Foundation Models
AI Agents
Software Dev & Coding

Meta on September 2, 2026 released Muse Spark 1.3, an update it says improves agentic and coding work while using fewer tokens and tool calls than Muse Spark 1.2. The model is live in Muse Code and the Meta Model API; previously available reasoning modes ship now, while a max-reasoning variant is held for additional safety testing.

September 2, 2026 · Meta

Muse Spark 1.3: more work, fewer tokens

Meta Superintelligence Labs shipped a coding- and agent-focused update that, in internal tests, used about 25% fewer tokens and 20% fewer tool calls than Muse Spark 1.2. Max reasoning is still in safety review.

~25%

fewer tokens vs 1.2 (Meta engineers)

~20%

fewer tool calls vs 1.2

AA Intelligence Index (xhigh)

$1.25 / $4.25

standard in / out per 1M tokens

DeepSWE v1.1 (company table)

Long-horizon agentic coding: Muse Spark 1.3 listed at 75.4 vs 1.2 at 55.0, edging Opus 5 at 74.0.

55.0

Spark 1.2

74.0

Opus 5 (max)

75.4

Spark 1.3

Token load vs predecessor

Same jobs, fewer tokens: a 4-block column (prior usage) beside a 3-block column (~25% less).

1.2 baseline

1.3 ~25% fewer

What users like

Cheaper completed jobs, less filler, and standard rates that undercut several frontier peers. Contributor SKU drops input to $0.10 per 1M if training on the traffic is acceptable.

Limits called out

Max reasoning is not public yet. Contributor has tighter RPM. Independent scores still trail on some browsing and instruction-following tests; first-token latency is high.

Where it runs

Live today in Muse Code and Meta Model API

macOS and Linux installer for Muse Code · 1M-token context · open weights and larger models remain on the published roadmap, undated · max reasoning after safety tests.

Continue reading

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