Updated: August 2026
Meta Muse Spark 1.1 is a multimodal agentic reasoning model released on July 9, 2026, by Meta’s Superintelligence Labs — the company’s first paid AI model available via a commercial developer API — priced at $1.25 per million input tokens and $4.25 per million output tokens, roughly 25% of what comparable models from Anthropic and OpenAI cost, according to Meta CEO Mark Zuckerberg.
The launch marks a deliberate strategic shift for Meta, which has historically distributed AI models for free via the open-source Llama family. Muse Spark 1.1 is closed-weight and sold as a subscription API, placing it in direct competition with Claude Sonnet 5 and OpenAI’s GPT-5.6 Sol, Terra, and Luna in the fast-growing agentic coding and tool-use market.
The API is available in the United States only at launch. New accounts receive $20 in free credits. The context window is 1 million tokens, and the API is OpenAI-compatible — developers can switch to Muse Spark by changing a base URL and API key, with no code rewrite required.
Published: July 9, 2026 · Source: Meta AI Blog
How Does Muse Spark 1.1 Pricing Compare to Claude Sonnet 5 and GPT-5.6?
Meta has priced Muse Spark 1.1 at $1.25 input / $4.25 output per million tokens — between 50% and 75% cheaper than its two main rivals on a per-token basis, according to verified pricing data from Anthropic, OpenAI, and Meta’s own announcement.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Context | OpenAI-compatible |
|---|---|---|---|---|
| Meta Muse Spark 1.1 | $1.25 | $4.25 | 1M tokens | Yes |
| Claude Sonnet 5 | $2.00 | $10.00 | 200K tokens | Partial |
| GPT-5.6 Terra | $2.50 | $15.00 | 128K tokens | Yes |
Source: Meta AI Blog (July 9, 2026), The Decoder (pricing comparison), OpenAI and Anthropic public pricing pages.
What Can Muse Spark 1.1 Do? Benchmarks and Capabilities
Muse Spark 1.1 is a multimodal reasoning model built for agentic tasks — workflows where the model plans, uses external tools, and executes multi-step actions. According to DevelopersDigest’s independent evaluation, it tops JobBench and MCP Atlas (the two standard tool-use benchmarks), meaning it outperforms Claude Sonnet 5 and GPT-5.6 Terra on agent coordination tasks specifically.
Pure coding benchmarks tell a different story. On SWE-Bench Pro and Terminal-Bench, Muse Spark 1.1 trails both Claude Opus 4.8 and GPT-5.5 — models priced at a significant premium. Developers whose primary use case is generating, reviewing, or debugging code in isolation will find the incumbents stronger. Developers building agents that orchestrate tools, APIs, and MCP servers will find Muse Spark competitive at a fraction of the cost.
The API supports zero-shot MCP (Model Context Protocol) server integration, which lets agents generalize to new tools without per-tool fine-tuning. Rate limits are 60 requests per minute and 2 million tokens per minute on the free tier, and 3,000 requests per minute and 4 million tokens per minute on paid plans.
Why Is Meta Charging for a Model for the First Time?
Meta’s Superintelligence Labs — rebranded from FAIR in early 2026 and now led by Alexandr Wang, former CEO of Scale AI — is operating as a separate commercial division. Muse Spark is that division’s first revenue-generating product, distinct from the Llama open-source research family which remains free and open-weight.
Zuckerberg posted on X for the first time in three years to announce Muse Spark 1.1, a signal of how much strategic weight Meta has assigned to the launch, according to Fortune’s reporting. The announcement positions Superintelligence Labs as a standalone business unit competing directly with Anthropic and OpenAI, not just a research organization releasing models for ecosystem goodwill.
The Llama bifurcation is deliberate: Meta appears to be running two parallel strategies. Open-source Llama models commoditize the general AI market and create developer goodwill; closed Muse Spark models extract direct revenue from the commercial developer segment that requires guaranteed SLAs, enterprise rate limits, and multimodal tool-use capability.
Who Can Use Muse Spark 1.1?
Muse Spark 1.1 is available only to US-based developers at launch, via the Meta AI developer portal at ai.meta.com. International availability has not been announced. The API uses the OpenAI API format, so any application already integrated with OpenAI’s SDK can switch by updating the base URL and API key — no additional SDK installation is required.
Meta’s Muse Spark pricing follows a pattern of rapid deflation in foundation model costs. Claude Sonnet 5 launched in May 2026 at $2/$10 per million tokens, itself a reduction from earlier Anthropic pricing. OpenAI introduced its multi-tier GPT-5.6 Sol, Terra, and Luna line in June 2026 with Terra at $2.50/$15 as the mid-range option. Muse Spark undercuts both by 50–75% on a per-token basis while claiming parity or better on agent-specific benchmarks.
August 2026 Update: Muse Spark 1.2 and Muse Glimmer Open Weights
Meta released Muse Spark 1.2 on August 5, 2026, as a closed-weight upgrade targeting complex software engineering tasks. The model scores 54 on the Artificial Analysis Intelligence Index, placing it level with Grok 4.5 and above Muse Spark 1.1 on reasoning benchmarks, according to Artificial Analysis’s independent evaluation.
On August 10, 2026, Meta reversed the proprietary strategy it adopted in April: Zuckerberg and Alexandr Wang announced on X that open weights for Muse Spark 1.2 will be released “soon” under a permissive license, without a confirmed date. The announcement accompanied the immediate open-weight release of Muse Glimmer, a separate 30-billion-parameter multimodal model released the same day under the Apache 2.0 license.
Muse Glimmer is a 30B-parameter causal language model that combines a large language model with a dedicated perception encoder, supports both natural language and visual inputs, and runs entirely offline on a single 24 GB consumer GPU. Meta describes it as “one of the highest-performing models of its size” and positions it as agent-focused, according to the Meta AI announcement on August 10, 2026.
Zuckerberg published a 6,500-word essay on August 10 alongside the releases, arguing that concentrated control of advanced AI by a small number of companies, institutions, or governments is the technology’s primary risk. He positioned open-source AI distribution as the structural solution: broadly distributed model weights prevent any single actor from monopolizing AI capability. He also stated that Meta’s independent board will have the power to approve model-release safety criteria and review whether releases meet those standards, according to reporting by Axios and the Philadelphia Inquirer.
Meta’s August pivot reverses a strategy shift that began with the April 2026 Muse Spark 1.0 launch, when the company moved the Muse family to a closed, API-only model for the first time in its AI history. The August 10 announcements position the Muse family as the largest American open-source models, competing directly with open-weight models from Mistral and Alibaba’s Qwen family. Sources: CNBC, Axios.
Frequently Asked Questions
What is Muse Spark 1.2?
Muse Spark 1.2 is Meta’s second major version of its agentic reasoning API model, released on August 5, 2026. It scores 54 on the Artificial Analysis Intelligence Index, level with Grok 4.5, and targets complex software engineering tasks. Meta announced on August 10, 2026 that open weights will be released “soon” under a permissive license.
What is Muse Glimmer?
Muse Glimmer is a 30-billion-parameter open-weight multimodal AI model released by Meta on August 10, 2026 under the Apache 2.0 license. It supports natural language and visual inputs, runs entirely on a single 24 GB consumer GPU, and is designed for agentic tasks. Meta describes it as one of the highest-performing open-weight models of its size.
Is Muse Spark 1.1 still available?
Yes. Muse Spark 1.1 remains available via the Meta AI developer API at $1.25 per million input tokens and $4.25 per million output tokens. Developers in the United States can access it at ai.meta.com. Muse Spark 1.2 is also available as a closed-weight API, with open weights announced for future release.
How does Muse Spark compare to Claude Sonnet 5 and GPT-5.6?
Muse Spark 1.1 and 1.2 are priced 50–75% below Claude Sonnet 5 ($2/$10 per 1M tokens) and GPT-5.6 Terra ($2.50/$15 per 1M tokens). On agent coordination benchmarks (JobBench, MCP Atlas), Muse Spark 1.1 outperforms both. On pure coding benchmarks (SWE-Bench Pro, Terminal-Bench), Claude Opus 4.8 and GPT-5.5 lead.
Why did Zuckerberg publish an open-source AI essay in August 2026?
Zuckerberg published a 6,500-word essay on August 10, 2026 arguing that concentrated AI control by a small number of companies or governments is the technology’s biggest risk. He used the essay to accompany the release of Muse Glimmer’s open weights and the announcement that Muse Spark 1.2 weights would follow, framing open-source distribution as a structural safeguard against monopolization.
Last updated:
