On July 9, 2026, Meta launched Muse Spark 1.1 as its first paid developer API model, entering a market long dominated by OpenAI and Anthropic. The pricing is aggressive: $1.25 per million input tokens and $4.25 per million output tokens—roughly one-quarter of Claude Opus 4.8 and GPT-5.5 rates. New accounts receive $20 in free credits. The model features a one-million-token context window with active compaction and is compatible with both OpenAI and Anthropic SDK formats.
The model arrives under the direction of Alexandr Wang, who heads Meta Superintelligence Labs. At the Bloomberg Tech Summit on June 4, Wang was candid about where Muse Spark sits: “The new Muse Spark model that we released is not at the tier of the leading frontier models… But we believe it’s a very exciting data point on the trajectory.” Asked when a frontier-tier model would arrive, Wang said: “We’re cooking it. We’re seeing very exciting and promising results in the process of training it right now.” He also explained the move to a closed model: “When the company launches a model in a product, we have a lot of ways to mitigate some of these risks. It’s much harder to do that when you open-source the model.”
Independent benchmarks offer a more granular picture. Vals AI ranked Muse Spark 1.1 fourth on its Vals Index, calling it “particularly fast and cost-effective” and reporting it as the fastest model in the top 10—running roughly three times faster than peers. The model also set new state-of-the-art scores on the MedScribe and TaxEval benchmarks, and led all models on the Harvey legal-agent benchmark with a 20% score, compared to 11% for Fable, 9% for Claude Opus 4.8, and 4% for GPT-5.5.
But the gap between Meta’s own reported benchmarks and independent testing has drawn developer scrutiny. Meta reported a Terminal-Bench 2.1 score of approximately 80, while Vals AI independently measured 69.29. “They are showing 80 score on terminal bench 2.1 but Vals AI showing 69.29. Now how to know which is right?” one developer asked on Reddit. The SWE-bench Pro score of 61.5 remains vendor-reported and is not independently verified. Meta has not published a model card, architecture details, or training data documentation—limiting external validation.
The pricing structure carries a hidden cost. Reasoning tokens are billed at full output rates, meaning complex agentic tasks that trigger extended reasoning chains could run significantly higher than the headline $4.25 figure suggests. For developers building multi-step agent workflows, the effective cost per task may be closer to Claude or GPT than the raw per-token comparison implies.
Muse Spark 1.1 is positioned as Meta’s strongest model for agentic tool use, not pure coding accuracy. It trails Claude Opus 4.8 and GPT-5.5 on raw coding benchmarks but leads on professional tool-use assessments like JobBench and scores 88.1 on MCP Atlas. The strategic question is whether Meta is buying market share with below-cost pricing or building a genuine competitive moat in the agentic layer. Wang’s “appetizer” framing suggests the latter: Muse Spark is the proof of concept, and the entrée—the frontier model—remains in training.