Meta's Muse Spark 1.3 Ranks 5th on LiveBench at $0.219 per Task — Cheapest Proprietary Path to the Top 5
Meta’s Muse Spark 1.3 scored 81.6 on LiveBench overall, landing fifth behind four Anthropic and OpenAI models, but its $0.219 cost per successful task makes it the cheapest proprietary route into the frontier top five by a wide margin.
LiveBench publishes cost-per-task alongside raw scores, giving a direct view of what performance costs in production. That column is where Muse Spark 1.3 separates itself from the competition.
The Cost-per-Task Gap
The top five on LiveBench as of September 23, 2026:
| Model | Overall | Cost/Task |
|---|---|---|
| Claude Fable 5.1 (Max Effort) | 83.4 | $1.212 |
| Claude 5.5 Opus Thinking (Max Effort) | 83.2 | — |
| Claude Fable 5 (Max Effort) | 83.0 | $1.439 |
| GPT-6 Astra (Max Effort) | 82.2 | $0.736 |
| Muse Spark 1.3 (xHigh Effort) | 81.6 | $0.219 |
At $0.219 per task, Muse Spark 1.3 costs roughly 18 cents for every dollar spent on Claude Fable 5.1, while scoring within 1.8 points overall. For throughput-heavy workloads where volume matters (batch coding, document processing, agentic pipelines), that differential compounds quickly.
Three models in the full top 15 cost less: DeepSeek V4.1 Flash at $0.029 per task (open weights, ranked sixth, 81.1 overall), Gemini 3.7 Flash High at $0.157 per task (ranked 12th), and Grok 4.6 at $0.207 per task (ranked 13th). Among the top 10, Muse Spark 1.3 is the most affordable proprietary option by a clear margin.
Where Muse Spark 1.3 Scores
Muse Spark 1.3’s subcategory breakdown matters as much as the headline number. Its strongest showing is reasoning (89.7), matching Claude Fable 5’s reasoning score exactly, though still 2 points behind Claude Fable 5.1’s 91.7. Its weakest is agentic coding (64.1), which still beats GPT-6 Astra (57.3) and GPT-5.6 Sol (56.2).
The model scored 95.9 on mathematics, 81.1 on coding, and 78.0 on instruction following. Its overall gap to the top three is narrow enough that Muse Spark 1.3 represents a real trade-off rather than a budget compromise.
What Muse Spark 1.3 Is
Meta released Muse Spark 1.3 on September 2, 2026, built by its Superintelligence Labs division under Chief AI Officer Alexandr Wang. The model targets agentic workflows: long-horizon task tracking, multi-workflow threading in a single context, and proactive gap identification when given open-ended objectives.
Meta reported that Muse Spark 1.3 uses roughly 25% fewer tokens and 20% fewer tool calls than Muse Spark 1.2. Wang described it to Bloomberg as competitive with Claude Fable 5.1 and better than GPT-5.6 Sol on coding tasks. Artificial Analysis placed it at 62 on its Intelligence Index, behind only Claude Fable 5.1 and Claude Opus 5 among currently tracked models.
The model is available through the Meta Model API at $1.25 per million input tokens and $4.25 per million output tokens at the standard tier. A “contributor” pricing tier at $0.10 input and $0.20 output per million is available for developers who agree to let Meta use queries to improve its models.
Meta’s Proprietary Turn
Muse Spark 1.3 is part of Meta’s deliberate move away from the Llama open-weights strategy. The company has shipped proprietary models under the Muse brand since April 2026, when it launched the first Muse Spark from a rebuilt AI stack. Weights for Muse Spark 1.2 were promised but have not been released. Wang gave no timeline for 1.3 weights.
Meta’s next model, internally called Watermelon, is a larger architecture. Wang said it will be “extremely competitive” but declined to give a release date.
The pattern is consistent across four Muse Spark releases in five months: aggressive iteration, benchmark claims against Anthropic and OpenAI at each step, and price as a competitive lever. Whether that pricing holds as Watermelon arrives is the open question.