Meta on September 2 released Muse Spark 1.3, the fourth iteration of its flagship Muse Spark model in five months, pushing the family further into frontier territory on coding, agentic tool use and long-horizon reasoning. Developers gained paid API access on September 4, with consumer rollout across Meta AI, Instagram and Facebook flagged as imminent.
Frontier-class scores
Muse Spark 1.3 (xhigh) posts 61 on the Artificial Analysis Intelligence Index, up four points from Muse Spark 1.2 in August and eight points from 1.1 in July. The gated "max" reasoning tier, still in limited preview for Meta partners, reaches 62 — trailing only Claude Fable 5.1 and Claude Opus 5. Meta says the model burns roughly 20% fewer tool calls and 25% fewer tokens than 1.2 while covering the same evaluation set.
Longer sessions, multimodal by default
The update targets a specific weakness in earlier Muse Spark releases: coherence over long, tool-heavy sessions. Muse Spark 1.3 ships with a one-million-token combined context window, accepts text, images and video as input, and is tuned to sustain multi-workflow agent runs in a single thread rather than restarting on every turn.
Pricing holds, safety gates max mode
Meta kept published rates flat versus 1.2 at $1.25 per million input tokens and $4.25 per million output tokens, undercutting most frontier competitors. The max reasoning mode is still gated pending additional safety evaluation. The launch follows Meta's earlier Muse Spark 1.2 and Muse Code drop, the open-weight Muse Glimmer 30B release and last month's cyber-red-team validation of Muse Spark 1.1 by Irregular.
The rapid cadence signals Meta Superintelligence Labs is now shipping on roughly a monthly rhythm, staying within striking distance of OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 as the September frontier race intensifies.
Reporting based on coverage from Meta AI Research, Bloomberg, Axios, Artificial Analysis, Winbuzzer and TechJournal.