Somebody changes a model string in a config file on Tuesday, the coding agent still compiles, and the standup note says the team is on Muse Spark now. What that person has signed up for is a second evaluation harness, because nobody switches off the incumbent while the replacement is still being trusted. In practice, when a cheaper frontier model lands, developers route the low-risk half of the traffic to the new vendor, keep the old one live, and pay two bills plus the labour of comparing them. The token count is the figure worth putting in a spreadsheet, more so than the index score. Wang told Bloomberg that developers pay no more for 1.3 than they did for 1.2, which launched in August [6], and that it gets through the same tasks on roughly 25% fewer tokens [7]. Hold the rate flat and cost per task falls to three quarters of what it was, so a fixed monthly budget buys about a third more work [20]. Both inputs come from Meta, and neither has been checked outside it. The saving also only reaches the invoice where tokens dominate the cost of a workload, which is not true of anything whose expensive part is a human reading the output. Artificial Analysis scored Muse Spark 1.3 (max) at 62 on its Intelligence Index, behind only Claude Fable 5.1 and Claude Opus 5, and that variant sits in limited preview for Meta's partners [4]. The number therefore describes something most buyers cannot call this week. SiliconANGLE's own caution is the right one to carry into a vendor meeting: models are uneven across tasks and published benchmarks are easy to game [12]. Four Muse Spark releases have shipped in five months [3], starting with a closed-source model in April [15], which averages out to a drop roughly every seven weeks [21]. Whoever owns your regression suite owns the cost of that cadence, since the model you certify in October won't be the one on sale in December. Wang also cited developers running trillions of tokens per week [9]. That figure measures consumption, not whether a pull request landed or how many had to be redone. The numbers that decide whether a third vendor stays are cost per completed task, the rework rate behind it, and how long the first working integration took. Two behaviour changes matter more than the score. The model can hold several workflows without a separate session for each and is better at retaining context across tasks [10]. It also asks for confirmation before every irreversible action [11], a safety feature that carries a real throughput cost: unattended agent runs now stop and wait for a person. For anyone with a self-hosting plan or a European filing, the licence is unsettled. Meta has not decided whether to publish 1.3's weights and still intends to release 1.2's [15]. Article 53 of the AI Act exempts genuinely free and open-source general-purpose models from the technical documentation owed to the AI Office and to downstream developers, but only where weights, architecture and usage information are all public with no non-commercial clause [16]. That relief stops at models classified as carrying systemic risk, which owe every obligation whatever the licence says [17]. A permissive licence on 1.3 would change Meta's paperwork; it would not remove it. The sort for Monday has two axes: token volume per task, and whether the output is reversible without a human in the way. High volume and reversible, such as test scaffolding or refactors that get reviewed anyway, is where 25% fewer tokens at a flat rate actually shows up, and where a wrong answer costs a rerun.
