Meta Enters AI Coding Race With $1.25/M Muse Spark 1.1

Meta just became a paid API competitor — and the pricing is aggressive enough to reshuffle the cost floor of AI coding. On July 9, 2026, Meta shipped Muse Spark 1.1, its first-ever paid model, at $1.25 per million input tokens and $4.25 per million output tokens through the new Meta Model API. That output price undercuts GPT-5.6 Sol ($30/M) by roughly 86% and Claude Opus 4.8 ($25/M) by 83%. For the first time, Meta is charging for a model — and it's charging far less than the incumbents.

Mark Zuckerberg broke a three-year silence on X/Twitter to announce it, calling Spark "a strong agentic and coding model at a very low price" and teasing that "more to come soon." The model comes from Meta's Superintelligence Labs division, led by AI chief Alexandr Wang — a deliberate departure from the traditional Meta AI Research organization behind Llama.

What happened

Muse Spark 1.1 is a multimodal reasoning model handling text, images, and audio in a single architecture, with a self-managed 1-million-token context window. Instead of keeping everything in active memory, the model decides what to retain, what to retrieve from earlier in a session, and what to compress. For long agentic runs, that matters because naive context-stuffing burns money on every step.

The architectural differentiator is native multi-agent orchestration: the same model can function as both a primary agent and a subagent, breaking large tasks into parallel subtasks and gathering results. It ships with built-in MCP (Model Context Protocol) support, custom skills, and direct computer control — patterns that require external frameworks when using most competitor models.

On benchmarks, the clearest public number is MCP Atlas, a test of scaled tool use: Muse Spark 1.1 scores 88.1, which Meta claims beats Claude Opus 4.8 and GPT-5.5 (both in the high 70s to low 80s). The original Muse Spark (April 2026) was openly weak on coding and long-horizon agentic work. Version 1.1 is engineered specifically to close those gaps.

But independent verification is thin. Meta claims the model runs competitively on its internal coding benchmark, framing it as GPT-5.5-class on agentic coding. Independent SWE-bench Verified scores were not published at launch, so the coding claims remain vendor-reported. Third-party independent testing places Muse Spark 1.1 in the same competitive cluster as Grok 4.5, Claude Opus 4.8, GPT-5.5, and GLM-5.2 — with Claude Fable 5 ahead of that group.

Why it matters

This isn't just another model launch — it's Meta's strategic pivot from open-weight to paid API, and it lands in the middle of an accelerating AI price war. The same week saw GPT-5.6 Sol/Terra/Luna go public and Grok 4.5 ship at $2/$6 (per xAI pricing; see /models/grok-4-5). A practical pattern is emerging: run most agentic steps on Muse Spark 1.1 (or Grok 4.5 at $2/$6), escalate only the hardest tasks to a flagship like Claude Opus 4.8 or GPT-5.6 Sol.

Meta spent years building goodwill with the open-weight Llama series. Muse Spark 1.1 launches as a paid API model — a direct challenge to OpenAI and Anthropic on their own business model, led by Alexandr Wang rather than the traditional open-source AI research arm. Zuckerberg's return to X after three years explicitly to announce this model signals how important Meta considers this market.

Pricing-wise, Muse Spark 1.1 sits above GPT-5 mini and Claude Haiku 4.5 but well below Claude Sonnet 4.6. Against the flagships, the gap is dramatic: output tokens are roughly 7x cheaper than GPT-5.6 Sol and 6x cheaper than Claude Opus 4.8. The question is whether enterprise developers will trust a first-generation paid API from a company whose AI brand until now has been "free and open."

What changes for you

US developers can evaluate Muse Spark 1.1 now. The Meta Model API is in public preview with $20 in free credits for new accounts, then pay-as-you-go. Consumers can access it in Thinking mode through the Meta AI app and meta.ai.

Don't commit production workloads yet. Until independent SWE-bench Verified scores land, the coding claims are vendor-reported. The independent competitive picture — placing Spark 1.1 in the same cluster as Grok 4.5 and Opus 4.8 rather than ahead of them — suggests the model is competitive, not dominant. Use the free credits to benchmark against your own workloads before making a switch.

For teams already running multi-model pipelines, Muse Spark 1.1 at $4.25/M output slots naturally into the budget tier — run routine agentic steps on Spark or Grok 4.5, escalate complex tasks to Opus 4.8 or GPT-5.6 Sol. The price gap is wide enough to meaningfully change cost-per-task for high-volume agentic pipelines.

FAQ

What is Muse Spark 1.1? Meta's first paid AI model, built by the Alexandr Wang-led Superintelligence Labs division. It's a multimodal reasoning model with a self-managed 1M-token context window, native multi-agent orchestration, and built-in MCP support — engineered specifically for agentic coding workloads rather than general-purpose chat.

How does the pricing compare to other frontier models? At $1.25/$4.25 per million tokens, Muse Spark 1.1 undercuts GPT-5.6 Sol's output price by ~86% and Claude Opus 4.8 by ~83%. It's priced above the budget tier (Haiku 4.5, GPT-5 mini) but well below mid-tier models like Sonnet 4.6.

Is Muse Spark 1.1 the strongest coding model available? No. Meta claims GPT-5.5-class agentic coding performance, and independent testing places it in the same competitive cluster as Grok 4.5, Opus 4.8, and GLM-5.2 — with Claude Fable 5 ahead. Independent SWE-bench Verified scores were not published at launch, so treat vendor claims as directional until third-party benchmarks land.