NCCapital

Meta’s recent progress with its AI models appears significant. Muse Spark 1.2 has reportedly surpassed Google’s offerings in quality for many use cases, a development that is particularly notable given the relatively short timeframe involved.

The current models still belong to the Muse Spark family. Meta has already indicated that larger and more capable systems, referred to as the Watermelon family, are forthcoming. Based on Muse Spark 1.2’s performance, these next models could reach the capability level of Anthropic’s Fable series.

After roughly a year of internal overhaul, Meta seems to have established solid scaling foundations for its AI lab. The company is now demonstrating strong product shipping velocity, with three model releases delivered in just four months.

Access to sufficient computing capacity is becoming a critical competitive factor. Recent reports of potential price increases for DeepSeek users highlight that even a strong model has limited impact if it cannot be widely served to customers.

Meta stands among the few companies with compute resources comparable to, or potentially exceeding, those of Anthropic and OpenAI. This infrastructure advantage could prove essential as demand for advanced AI models continues to grow.

Meta has released Muse Spark 1.2, its third model update in four months. The new version scores 54 on the Artificial Analysis Intelligence Index, delivering meaningful gains in agentic knowledge work and placing Meta in a tie for third among U.S. AI labs alongside SpaceXAI.

Muse Spark 1.2 improved three points over Muse Spark 1.1 and eleven points over the April release of Muse Spark 1.0. It now sits effectively tied with GPT-5.5 and Grok 4.5, while remaining behind frontier systems such as Claude Opus 5, Claude Fable 5, GPT-5.6 Sol, and Kimi K3.

The model shows particular progress on agentic knowledge work, an area previously identified as a relative weakness. Its GDPval-AA v2 Elo rose 260 points to 1631, ranking fifth among all models benchmarked and surpassing Claude Opus 4.8. Terminal-Bench 2.1 and Tau3-Bench Banking also posted modest gains.

Muse Spark 1.2 ranks among the more cost-efficient models at its intelligence level. At Meta’s unchanged pricing of $1.25/$4.25 per million tokens, it costs roughly $0.40 per Intelligence Index task—competitive with Grok 4.5 and GPT-5.6 Sol and cheaper than several higher-priced alternatives.

Other metrics show a higher abstention rate that reduced hallucinations but also slightly lowered accuracy, while scientific reasoning results remained largely stable. The model retains a 1-million-token context window and is available through Meta’s first-party API at launch.