Meta Muse Spark 1.3 ships: ~25% fewer tokens per task, same price
It also cuts tool calls ~20%, lands in Muse Code and the API at unchanged prices, and Meta calls it frontier-competitive. Plus a new European coder via API.
Copy markdown
Fewer tokens, same bill
Spark 1.3 uses ~25% fewer tokens and ~20% fewer tool calls than 1.2 on coding and agentic tasks, and finishes in fewer turns — while pricing holds at 1.2's rate card of $1.25/M input and $4.25/M output. For anyone running Muse agents at any volume, that's a direct cost cut with nothing to migrate.
Live today in Muse Code and the API
It's available now through Muse Code at dev.meta.ai and the Meta Model API, keeping the 1M-token context window. No waitlist — you point your existing calls at the new model string and go.
Meta says it's frontier-competitive
Meta's scorecard puts 1.3 head-to-head with GPT-5.6 Sol (max) and Opus 5 (max) across agent, coding, instruction-following, and long-context evals, and AI chief Alexandr Wang calls it 'very competitive with frontier models.' A 'max reasoning' mode is being held back pending more safety testing.
Tuned for long, messy runs
The upgrade is aimed at where agents actually break: sustaining longer tasks, juggling multiple workflows in one thread, proactively patching gaps in its own plan, and using better judgment before irreversible actions. That's more useful day-to-day than a benchmark point.
Elsewhere: Quasar 438B, a European coder via API
Multiverse Computing shipped Quasar 438B on its CompactifAI API (dashboard.compactif.ai) — no open weights, just sign up and send requests. It scores 69.3 on Terminal-Bench v2.1 and 43 on Artificial Analysis' Intelligence Index (top among European models), does English and Spanish, and returns 500 tokens in ~15s.