AI Pulse · Frontier Labs Watch

Meta has released Muse Spark 1.3 broadly, priced at $1.25 per million input tokens. The company says the available version uses 20% fewer tool calls and 25% fewer tokens than its predecessor. For a business running coding agents or automated development workflows, that matters because tool calls and token use are where costs can climb quickly. A model that completes the same work with less back-and-forth could lower the cost of routine engineering tasks.

The more interesting detail is what Meta has not made widely available. Its stronger Muse Spark 1.3 Max variant is still in partner-only preview. Meta says that version ranks second on Artificial Analysis’s Coding Agent Index, behind only Claude Opus 5. That gives Meta a credible claim that it has closed the gap with Anthropic and OpenAI, but outside companies cannot yet broadly test, benchmark, or deploy the model that supports the claim.

Treat this as a signal, not a procurement decision. The public model may be worth testing where coding-agent efficiency matters, especially if your team already measures cost per completed task rather than just model price. But don’t assume the headline performance applies to what you can actually buy today. Ask vendors which model variant powers the product, what access tier you receive, and whether their reported results come from public or preview models. This is the kind of thing we build into an AI command centre, tracking the gap between lab claims, available products, and real operating value.