Meta's Muse Spark 1.3 Reaches the Frontier: 75.4% on DeepSWE, 1M Context, and a Data-for-Discount Endpoint

Meta has released Muse Spark 1.3, its most powerful AI model to date, with the company's chief AI officer saying its capabilities are edging closer to OpenAI and Anthropic's frontier — a claim the independent numbers largely support.

The numbers

Muse Spark 1.3 lands at #6 of 636 models on the Artificial Analysis Intelligence Index, with a 1 million-token context window and text, image and video input. The headline results are Meta's biggest generational jump on agentic work:

• 75.4% on DeepSWE 1.1 end-to-end agentic software engineering
• 88.8% on Terminal-Bench 2.1
• 59.4% on SWEAtlas CodeBase QnA
• 98.5% on long-context retrieval (MRCR)

Efficiency improved alongside raw capability: Meta's engineers measured the model completing coding work with roughly 20 percent fewer tool calls and 25 percent fewer tokens than Muse Spark 1.2 — the metric that increasingly determines real-world agent economics.

The pricing experiment

The release's most provocative feature is commercial, not technical. The standard API endpoint costs $1.25/$4.25 per million tokens. But a "contributor" endpoint runs 10-20x cheaper — if you let Meta train on your data.

No other frontier lab has priced the data-for-compute trade this explicitly. For hobbyists and open-ended research the discount is compelling; for enterprises it effectively prices data governance: choosing the cheap tier means feeding your codebase and workflows into Meta's next training run. Expect the structure to be copied — or regulated — quickly.

Meta's long road back

A year ago Meta's superintelligence reorganization looked chaotic: leadership churn, a shelved Behemoth, and Mark Zuckerberg's own admission that AI agents were coming along slower than hoped. Muse Spark 1.3 is the first release that makes the retooled lab look like a frontier contender again — sitting a tier below GPT-6 Astra and the newest Claude and Gemini flagships on raw intelligence, but competitive where enterprise money actually flows: agentic coding, long context and cost.

The frontier is now a four-way race with a crowd of fast followers — and with Chinese labs like Z.ai and Moonshot compressing prices from below, Meta's aggressive pricing may matter more than its benchmark position.