Wait Which Model?
← Back to directory
Meta AI

Muse Spark 1.3

Released Sep 2, 2026 · knowledge cutoff unpublished

Status
Frontier
Location
United States
Modality
Multimodal
Context window
1M
Max output
—
Speed (xhigh effort)
179 tok/s · 42s to first answer token
Price ($/MTok in / out)
$1.25 / $4.25
Cost per task (xhigh effort)
$0.55
Open weights
No
Benchmarks5 of 10 reported

GPQA Diamond

94%

Terminal-Bench 2.1

85%

Humanity's Last Exam

47%

LMArena Elo

1493

GDPval-AA v2

1709

MMLU-Proretired

—

SWE-bench Verified

—

SWE-bench Pro

—

AIME (latest)retired

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Trained to stop and ask — checks in before irreversible actions and surfaces conflicting inputs rather than picking one silently
  • Terser in agent loops than 1.2: roughly a fifth fewer tool calls and a quarter fewer tokens on the same coding work
  • The million-token window is genuinely usable — near-ceiling recall on Meta's own 512K–1M retrieval evals rather than degrading past the halfway mark
  • Streams fast for a reasoning model, so the cost of the long thinking phase is paid up front rather than throughout the answer

Weaknesses

  • Long silence before the first answer token — the reasoning phase runs well over half a minute at the shipping effort setting
  • Answer verbosity climbed sharply over 1.2, so the unchanged per-token price does not mean unchanged bills
  • The headline agentic numbers come from a max reasoning mode that was still in limited preview at launch — what developers can call is the weaker xhigh configuration

Figures here describe the generally available xhigh reasoning configuration; a stronger `max` variant was still in safety testing at launch and only Artificial Analysis' partner preview had access. Terminal-Bench 2.1 conflicts: Artificial Analysis measured 85% (xhigh) and 86% (max) in its own harness, while Meta reports 88.8% for max — the AA xhigh figure is recorded as the one matching the shipping model. GPQA Diamond, HLE and GDPval-AA v2 are Artificial Analysis measurements of xhigh, not lab-reported. Meta lists a Contributor tier at $0.10/$0.20 per M tokens in exchange for training rights, and cached input at $0.15/M. Meta's roadmap promises a future 'Muse Spark open weights release' but 1.3 itself is closed. LMArena Elo 1493 is the "muse-spark-1.3-max" text-arena listing (arena.ai, 2026-09-13 update, rank 8).

For developers

API model strings

  • metamuse-spark-1.3
  • openroutermeta/muse-spark-1.3

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor

News