Meta launches Muse Spark 1.1, an aggressively priced AI coding model to challenge Anthropic and OpenAI

Meta released Muse Spark 1.1 on July 9, a multimodal AI model built specifically for coding and multi-step agentic tasks, positioning it as a direct challenger to Anthropic's Claude and OpenAI's GPT models in the fast-growing agentic coding market. AI chief Alexandr Wang called it the company's "strongest model for agentic and coding work yet," as first reported by SiliconANGLE.

The launch matters because it marks Meta's clearest pivot yet from its open-source Llama strategy toward proprietary model monetization. Muse Spark 1.1 follows an initial version released in April 2026 that was limited to select partners; this update opens public preview access through the Meta AI app and a developer API waitlist, with a fully open-source variant reportedly still in development on an unconfirmed timeline.

On capability, Muse Spark 1.1 supports a 1-million-token context window and introduces a context compaction mechanism that compresses data generated across long agent workflows while preserving the ability to retrieve information from much earlier in a task. Meta also highlighted adaptive planning — the model can detect mid-task developments and revise its plan rather than sticking to a fixed initial approach, a common failure point for earlier agentic coding tools.

On benchmarks, Meta reported a score of 72.2 on its internal Vibe Code Bench v1.1, more than 50 points above its previous flagship model, and roughly 18% improvement on SWE-Atlas Codebase QnA testing. Independent evaluations put Muse Spark 1.1 at 61.5 on SWE-Bench Pro, trailing Claude Opus 4.8's 69.2 but ahead of GPT-5.5's 58.6 — a solid but not category-leading position across most third-party coding benchmarks.

Pricing is where Meta is making its most aggressive move: $1.25 per million input tokens and $4.25 per million output tokens, with new API accounts starting with $20 in free credits. Wang described the strategy explicitly as pricing that "scales with immense consumption usage," undercutting typical Anthropic and OpenAI rates for comparable agentic-coding tiers. For engineering teams evaluating agentic coding tools, that price gap — even without a clear benchmark lead — is likely to matter more for adoption at scale than marginal accuracy differences on any single evaluation.