CONFIRMED

Meta stopped treating Muse Spark as a lab experiment. On September 2, 2026 Superintelligence Labs put Muse Spark 1.3 live in Muse Code and the Meta Model API, and Mark Zuckerberg called it the biggest jump the house has made on coding and agents so far. His line, frontier performance almost too cheap to meter, is marketing. The operational fact is narrower: the muse-spark-1.3 ID exists, context stays at 1 million tokens, standard pricing copied 1.2, and max reasoning has not shipped.

Five months, four versions, Llama on the bench

The family was born in April. 1.1 landed in July, 1.2 on August 5 with Muse Code, 1.3 less than a month later. That is not the old Llama cadence. It is a lab that decided to sell a closed model while promising open weights soon and a watermelon in Zuckerberg's next post. Alexandr Wang told Axios that 1.3 paves the way for personal agents that work around the clock. That is company intent, not a product in the Instagram feed. 1.3 enters first through the coding harness and the API. Anyone waiting for it in WhatsApp is still in line.

What changes in the agent bill

Meta Model API docs list muse-spark-1.3 on the standard tier and muse-spark-1.3-contributor on the discount. OpenRouter quoted $1.25 input and $4.25 output per million tokens, the same card as 1.2. The official pricing page still names 1.1 and 1.2 on standard and 1.2-contributor on the discount. Treat 1.3 as an ID upgrade, not a new rate card, until Meta rewrites the table. Contributor remains the real deal: $0.10 and $0.20, with traffic used to train the next model. Meta engineers say 1.3 uses about 20% fewer tool calls and 25% fewer tokens than 1.2. If that number holds outside the house, the invoice falls even with nominal prices frozen. If it does not, too cheap to meter stays a slogan.

The house table is not a ranking

Meta published its own chart with 1.3 at max against Opus 5 and GPT-5.6 Sol. On their numbers, the model edges Opus on DeepSWE v1.1, 75.4 versus 74.0, ties Sol on Terminal-Bench 2.1 at 88.8, and blows out long MRCR at 98.5 and 98.1 versus Sol's 91.5 and 73.8. On agentic search and instruction following, Sol still leads. GDPVal and JobBench sit in the middle, near Opus, not above it. That changes the reading only if you accept Meta's harness. Artificial Analysis has not published a 1.3 index. BenchLM, as of Tuesday, still lists 1.2 as the house's best public model. Max reasoning, the mode the table uses, is in safety testing. Comparing 1.3 max with Sol max before that mode exists in the API is comparing a slide with an endpoint.

MaxAssistant reading

Confirmed: Meta shipped 1.3, the ID is in the API, 1.2 is no longer the tip of the family. Hypothesis: Wang wants a personal agent cheap enough to live inside the app, and 1.3 is the engine, not the product. Feeling: test now in Muse Code and on the standard ID, with a token cap. Do not move production onto contributor if client data cannot train Meta. Do not buy a win over Opus or Sol until AA measures it. Open weights and the watermelon remain a CEO tease, not a model card.

Sources

Meta Superintelligence Labs, Introducing Muse Spark 1.3: https://research.meta.ai/blog/introducing-muse-spark-1-3 | Meta, Model API overview: https://ai.developer.meta.com/docs/overview.md | OpenRouter, Muse Spark 1.3: https://openrouter.ai/meta/muse-spark-1.3 | Prashant Ratanchandani, Muse Spark 1.3 in the API and harness, quoting Mark Zuckerberg: https://x.com/pratanchandani/status/2095239735015096444 | Axios, Meta debuts Muse Spark 1.3 as personal agent work continues: https://www.axios.com/2026/09/02/meta-debuts-muse-spark-13-as-personal-agent-work-continues