Muse Spark 1.3 goes live as Zuckerberg promises open-weights release 'soon'

Meta CEO Mark Zuckerberg promised on Wednesday that an open-weights version of the company's flagship Muse Spark model will arrive 'soon,' but the thing developers can touch today is Muse Spark 1.3, now live on Meta's API service and Muse Code CLI. The update is the latest in a string of refinements Meta has shipped since Muse Spark launched in April, and it is designed to fix the model's most annoying habit: wasting tokens and rambling into dead ends.

Version 1.3 is built for the agent and coding-assistant workloads that have become the industry's main battleground, and it lands in the same week as OpenClaw 2.0. According to Meta, the model now seeks human counsel more frequently when the correct path is unclear, asks clarifying questions on ambiguous prompts, invokes help from the user when stuck, and confirms before taking consequential actions. That behavioral change has a direct cost benefit: fewer turns and fewer tokens per task, which should make an already relatively affordable frontier model even less expensive to run.

For enterprise buyers, the timing matters. The AI agent market is shifting from demo-ready chatbots to systems that can be trusted with multi-step, long-horizon work, and the biggest complaint from builders is that models burn credits while confidently marching in the wrong direction. Meta's move directly targets that complaint. Teaching a model to say 'I need more information' instead of guessing is a small change in behavior but a large change in operational cost, especially for teams running thousands of agentic loops per day.

The competitive picture is tightening as well. Independent benchmarking by Artificial Analysis previously showed Muse Spark 1.2 performing on par with GPT 5.6 Terra and Z.AI's GLM 5.3 Flash. The benchmark outfit's latest results largely back Meta's claims for 1.3, showing a 4 point jump in overall intelligence that puts it in a dead heat with GPT 5.6 Sol, Claude Opus 5, and Grok 4.6 High. That is a crowded podium, but Meta is competing on price as much as raw score. The contributor tier, which trades lower rate limits and permission for Meta to use prompts for training data, offers steep discounts down to $0.002 per million cached input tokens, $0.10 per million input tokens, and $0.20 per million output tokens.

The pricing strategy is not subtle. Meta is using cheap inference to pull developers into its ecosystem, then dangling an open-weights release to keep them there. Open weights appeal to companies that want to run models on their own infrastructure, avoid per-token fees at scale, or retain full control over data. Zuckerberg's 'soon' leaves the exact timing vague, but the signal is clear: Meta wants to be the default choice for builders who are tired of paying premium prices to closed frontier labs.

For investors, the cadence of releases is doing some heavy lifting. Since launching Muse Spark in April, Meta has shipped several refinements to keep pressure on the competition and convince investors that its rampant capital spending is not for naught. Each incremental benchmark gain and price cut reinforces the story that the company can convert infrastructure spending into a durable, widely used AI platform. The risk is that 'soon' stretches into months, giving rivals room to match the open-weights move or undercut the pricing.

The strategic stakes extend beyond Meta. Every enterprise that standardizes on a model family is effectively making a bet on a vendor's roadmap, its pricing discipline, and its willingness to cede control through open weights. Muse Spark 1.3's improvements to token efficiency and human-in-the-loop behavior are exactly the kind of features that make a model easier to deploy in production, and the arrival of OpenClaw 2.0 the same week is a reminder that the open-source ecosystem is not standing still. For CTOs and AI leads, the takeaway is straightforward: the frontier is getting cheaper, more honest about its limits, and much harder to ignore.