Muse Spark 1.1
Meta’s new Muse Spark 1.1 model API is drawing attention for aggressive pricing and strong tool-calling and coding benchmarks that appear competitive with top models from OpenAI, Anthropic, and xAI. Commenters debate whether the quality justifies the cost, question Meta’s benchmarking practices and lack of transparency around data retention, and lament that the weights are closed despite Meta’s prior role in open-source AI. Many see the release as part of a broader trend of intensifying competition—especially from U.S. and Chinese labs—that is rapidly driving down token prices while reshaping expectations for software development and automation.
Pricing and Economic Positioning
- Many see Muse Spark 1.1 pricing as aggressively low vs other “frontier” models (e.g., Grok 4.5, Claude, GLM 5.2), especially cached input at $0.15/Mtok.
- Others argue it’s still expensive compared to Chinese models and open-weight options; some feel $10+ per 1M output tokens is hard to justify for consumers.
- Several comments frame this as a response to competitive pressure from Chinese labs and DeepSeek/GLM pricing.
- For professional use (coding, contracts), higher prices are seen as acceptable relative to human labor; for entertainment, less so.
Model Quality and Benchmarks
- Early testers report quality “below Sonnet” and not impressive on DeepSWE; some say it trades blows with GPT 5.5 / Opus 4.8 on Meta’s own charts.
- Tool calling and terminal/agent performance are highlighted as strengths; coding and multimodal are “pretty good,” with debate on how much high tool-call success matters in practice.
- Significant controversy over Terminal-Bench 2.1 results: critics say Meta exceeded CPU/RAM limits, effectively “cheating”; defenders argue resource limits are recommendations, impact may be small, and infra issues complicate strict adherence.
- Broader skepticism toward “trust-me” benchmarks and selective metric reporting; calls for independent evaluation.
Trust, Privacy, and Data Retention
- Multiple commenters say they won’t use Meta due to past privacy issues and lack of explicit data-retention policy for the paid API.
- Some note that not all LLM use involves personal data, but unease remains, especially vs open-weight models.
Open vs Closed Models
- Disappointment that Muse is not open-weights; seen as a strategic shift away from Meta’s earlier open-source leadership.
- Several argue that releasing strong open-weight models (and even training data, though viewed as legally risky) would let Meta “commoditize” frontier models and pressure competitors’ revenue.
Ecosystem, Access, and Competition
- Model is not yet on OpenRouter, making evaluation harder; people want a single-key aggregator to test it.
- Region restrictions (e.g., Canada, Argentina, Vietnam) frustrate developers.
- Overall sentiment: more competition (Meta, xAI, Chinese labs) is good for prices and diversity, but many remain wary of Meta’s brand and the closed nature of the release.