Meta's Muse Spark 1.3 briefly outpriced Gemini

Meta pushed its Muse Spark 1.3 reasoning model live on September 3, alongside Muse Code's official exit from beta and new subscription plans starting at $5 a month. Meta's own release claims the model's biggest jump yet on coding and agentic tasks, citing a 75.4% score on the DeepSWE coding benchmark, though the comparison figures against rival models come from Meta's self-described "best-effort" testing.

Artificial Analysis's independent numbers back the release. The publicly available Spark 1.3 xhigh scores 61 on the firm's Intelligence Index, tied with GPT-5.6 Sol max and Claude Opus 5 high, at roughly $0.55 per Intelligence Index task, the lowest cost of any model scoring 59 or above. That's a jump from Spark 1.2's $0.40 per task, driven by the newer model consuming about 57% more input tokens on agentic evaluations.

Timing amplified the release. Hours earlier, Google shipped Gemini 3.8 Flash at $0.75 per million input tokens and briefly held the top spot on Artificial Analysis's intelligence-versus-cost frontier with a 59 score at $0.58 per task. Spark 1.3 displaced it within hours.

Muse Spark's open-weight release, letting developers run the model outside Meta's hosted API, is still pending. Meta hasn't published the license terms that will govern how freely it can be modified or redistributed once available.