Mark Zuckerberg announced that Muse Spark 1.3 is rolling out now, with the model landing simultaneously in Muse Code and Meta's model API. He described the update as "the largest upgrade yet in this series for coding and agentic work," and said the cost of frontier-level performance has dropped to the point where it's barely worth counting.
The benchmarks tell half the story. On DeepSWE v1.1, Muse Spark 1.3 scored 75.4, ahead of GPT-5.6 Sol's 73.0 and Claude Opus 5's 74.0. It posted 88.8 on Terminal-Bench 2.1 and 59.4 on SWEAtlas CodeBase QnA. The two long-context tests came close to a perfect score, with 98.5 in the 256K–512K range of MRCR and 98.1 in the 512K–1M range. Compared with the previous 1.2 release, tool calls on the same coding tasks dropped by about a fifth, and token consumption fell by roughly a quarter. For anyone paying by the token, that number matters more than the benchmark score.
The cheap tier settles its bill in data
The line about cost being "barely worth counting" needs unpacking. Meta set two price tiers for Muse Spark 1.3: the standard tier runs $1.25 per million input tokens and $4.25 per million output tokens, while the contributor tier runs $0.10 per million input tokens and $0.20 per million output tokens. That's a 12.5x gap on input and a 21x gap on output.
The price difference comes with a condition spelled out in the terms: use the contributor tier, and your input and output get used by Meta to improve its products. That arrangement has been running since 1.2, and 1.3 simply carries it forward. For individual developers and small teams, a ten-cent input price is close enough to free. For companies with confidentiality requirements around their code, this tier is essentially off-limits — they're stuck with the $1.25 standard rate, which isn't especially cheap next to comparable models. The "barely worth counting" line is really about that first tier.
Ahead in coding, behind in agentic work
On Artificial Analysis's Intelligence Index, Muse Spark 1.3 (xhigh mode) scores 61, tying GPT-5.6 Sol (max) and Grok 4.6 (high), while Claude Opus 5 (max) stays ahead at 63.
The gap widens on agentic benchmarks. On GDPVal-AA v2, Muse 1.3 scores 1754 against Claude Opus 5's 1824; on OSWorld 2.0 it's 66.9 versus 68.3; on AutomationBench, 49.4 versus 50.3. Muse Spark 1.3 loses on all three, by narrow margins each time, but the direction is consistent. Zuckerberg's recurring line on earnings calls about personal agents working for you around the clock still looks like a way off, going by these numbers.
Two shoes yet to drop
A higher reasoning mode is waiting on safety testing before release, and an open-weight version is also planned — neither has a timeline. This 1.3 release added hardening against adversarial inputs and prompt injection, which is expected given how much the previous open-weight model got picked apart and repurposed after launch.
A rough estimate: if contributor-tier pricing holds, a coding agent chewing through five million input tokens a day would cost roughly $15 a month. That figure has nothing to do with technical sophistication, but it's what determines how many people fold Muse Spark into their daily workflow — and the code those people write becomes the training material for the next version.
Sources: Meta model API pricing page, Artificial Analysis Index, CocoLoop, OpenRouter model page; the two API price tiers and DeepSWE, MRCR, and Terminal-Bench benchmark figures were cross-checked line by line.