Muse Spark 1.1: Meta’s Paid AI API, Price And Benchmarks

Meta announced something it never announced before the price of its model. On July 9, 2026, Meta Super intelligence Labs announced Muse Spark 1.1 multimodal reasoning model based on agentic work and published it using a completely new and metered Meta Model API.

Meta Muse Spark 1.1 – Why Meta Charges for AI?

For many years Meta’s strategy was based on providing their models free of charge. Llama was a completely open weight free to download and free to fine tune model. Muse Spark 1.1 breaks that trend completely and provides a paid for product via a metered Meta Model API. Meta announces the public preview for US based developers with a wait list for new users.

What makes Muse Spark 1.1 unique is that Meta doesn’t try to beat its rivals on the benchmark leaderboard this time. Instead, the goal is to win in terms of price and distribution using its own platforms like WhatsApp, Instagram, Facebook and Ray-Ban glasses to attract developers. The new model is going to replace Llama powered chatbots in the same consumer platforms in the future.

Muse Spark 1.1 API Pricing & Cost

Muse Spark 1.1 costs $1.25 per million input tokens and $4.25 per million output tokens with $20 in free credits for new users. In comparison to Claude Opus 4.8 ($5 input / $25 output) and GPT-5.5 ($5 input / $30 output) the cost of the output of the new Meta model is cheaper than the rivals ones by a wide margin of about ($1.25 vs $5).

The problem is that agentic workloads are output heavy and generating a few million of output tokens in busy sessions may become expensive ($4.25 per million tokens). That is why the free credits are provided for testing but not for production use.

Meta Spark API is claimed to be OpenAI compatible with support of the structured outputs and parallel tool calling which decreases the switching cost for those developers who are building apps with GPT style SDKs.

Muse Spark 1.1 Benchmarks – Strong on Agents, Gap on Coding

Here comes the point where the approach becomes really interesting instead of cheap. Aggregated independently from Meta’s data, Muse Spark 1.1 outperforms other models in four out of twelve benchmarks MCP Atlas, JobBench, Humanity’s Last Exam with tools and Finance Agent v2 while Claude Opus 4.8 wins five and GPT-5.5 wins three. On a specific benchmark where Meta bets on a scaled tool Muse Spark 1.1 wins with 88.1 compared to 82.2 and 75.3 of Claude Opus 4.8 and GPT-5.5 respectively.

On pure coding, the results are reversed. On SWE-Bench Pro, Opus 4.8 wins with 69.2 while Muse Spark 1.1 loses with 61.5. This benchmark is even controversial and disputed by OpenAI itself which should be considered if you treat these numbers as gospel.

Muse Spark 1.1 is built to be an orchestrator in the first place. It can act as a lead agent which plans and delegates tasks to parallel subagents or as a sub agent which sticks to its mission and escalates if necessary. Also, it manages the million token context window actively compressing and retrieving previous results. This explains why Muse Spark 1.1 dominates tool orchestration benchmarks but lags behind on single shot coding accuracy as Meta optimized for coordination of tools and agents rather than becoming the single smartest coder.

The independent verification of Muse Spark 1.1’s performance is scarce. Artificial Analysis had rated the original April Muse Spark with the Intelligence Index of 43 which was placed mid pack behind Gemini 3.1 Pro, GPT-5.4 and Claude Opus 4.6 and as of now, Muse Spark 1.1 has not been benchmarked yet. Everything above this line is claimed by Meta.

Muse Spark 1.1 for Coding & Multimodal Tasks

Meta claims that the new model is capable of handling the coding tasks in the real world including diagnosis of bugs in big codebases, shipping features into the enterprise level systems, multi file migration and agentic coding capabilities like planning mode and sub agent delegation across various harnesses such as OpenCode. The reactions from the early partners were mostly positive: the CEO of Replit appreciated its coding abilities and compatibility with OpenAI and the CEO of Cline declared it to be viable for coding workloads due to its price alone.

On computer use, Meta provides a flagship demo which includes a “dinner party” agent which modifies its order mid session, if it gets new information and decides whether to script the action or click through the interface directly.

Meta Model API Access 

There is no access to Muse Spark 1.1 from the EU as it is a US only preview and the rate card, endpoints and model identifiers are all subject to change until the end of the public preview. Also, Meta ran the model through its Advanced AI Scaling Framework in terms of chemical/biological, cybersecurity and loss of control risk categories stating that it operates within safe margins. However, this is also a self reported result.

TheTweaks Verdict About Meta’s Muse Spark 1.1

Muse Spark 1.1 is not trying to become the smartest AI model on the market but the cheapest credible one which will be able to handle the agents that call tools, coordinate sub agents and run for a long time. If your workload is orchestration heavy (multi step agents, MCP tool chains, long running tasks) then Muse Spark 1.1 is hard to overlook due to its price to capabilities ratio and 88.1 MCP Atlas score. 

If your workload is single shot and hard coding problems Opus 4.8 and GPT-5.5 still have the edge. The bigger message here is that Meta charges for API access for the first time with the support of the consumer distribution which nobody else has. Every other lab pricing just got more pressure, use launch week numbers as a snapshot, not gospel verify current pricing and benchmarks before committing production traffic.