Muse Spark 1.3: Meta Learns to Stop Wasting Tokens
Meta released Muse Spark 1.3 on September 2, calling it their biggest jump in coding and agentic capability yet. It's live in Muse Code and the API now, rolling out to Instagram, Facebook, and Meta AI next, priced the same as 1.2. Alexandr Wang told Axios it brings Meta close to GPT-5.6-class performance; Zuckerberg says open-weight versions are coming.
The stat Meta chose to lead with is the tell: 1.3 needs roughly 20% fewer tool calls and 25% fewer tokens than 1.2 for the same work. Not smarter, cheaper to run. When labs start advertising efficiency per task instead of benchmark deltas, it means the buyers have changed β this is a pitch to people running agent fleets on a budget, not people reading leaderboards. Same-day context: Anthropic just cut cache-read prices 75% and Google shipped a promotional-priced Flash. All three frontier labs are suddenly competing on the cost of an agent loop.
Trained on more long-horizon coding tasks, same price as its predecessor, open weights promised. Meta spent the spring being mocked for the Muse deal-unwind mess; two point-releases later the line is genuinely competitive on agent work. HN gave the launch 250+ points.
Details: https://developer.meta.com/ai/models/muse-spark/
← Back to all articles
The stat Meta chose to lead with is the tell: 1.3 needs roughly 20% fewer tool calls and 25% fewer tokens than 1.2 for the same work. Not smarter, cheaper to run. When labs start advertising efficiency per task instead of benchmark deltas, it means the buyers have changed β this is a pitch to people running agent fleets on a budget, not people reading leaderboards. Same-day context: Anthropic just cut cache-read prices 75% and Google shipped a promotional-priced Flash. All three frontier labs are suddenly competing on the cost of an agent loop.
Trained on more long-horizon coding tasks, same price as its predecessor, open weights promised. Meta spent the spring being mocked for the Muse deal-unwind mess; two point-releases later the line is genuinely competitive on agent work. HN gave the launch 250+ points.
Details: https://developer.meta.com/ai/models/muse-spark/
Comments