Muse Spark 1.3 Review

Meta spent two years watching the frontier from behind, and this is the first release that lands on it. Published on September 2, Muse Spark 1.3 is a hosted multimodal model aimed squarely at long horizon work: a million tokens of context, strong recall across all of it, and a design meant to hold several threads of a job open at once instead of finishing one prompt and stopping.

The independent intelligence index that every lab quietly watches puts the widely available configuration level with the current tops from OpenAI and xAI, a place no Meta model has reached before, and the price is a fraction of theirs. Charges start at ten cents per million input tokens on the contributor tier, and a limited time contributor offer is the only route to trying it for nothing.

The asterisks are large. Meta's best published results come from a maximum reasoning configuration that was still finishing safety testing at launch and that developers cannot broadly use, a gap the trade press flagged within a day. The coding claim is shakier still: the benchmark Meta leads is one the model does not appear on publicly, and outside testers running ordinary build tasks describe output well short of the ranking, with at least one suite placing this point release under the version it replaces.

The cheap tier is also cheap for a reason, which the pricing page does not lead with: Meta trains on the prompts and completions that pass through it, and anyone routing client work through the ten cent tier is paying the difference in data. What survives is a genuinely capable and genuinely cheap workhorse for agent loops and bulk reasoning, plus a launch that got ahead of the engineering. Buy it for the context window and the bill, not for the chart.