Meta Muse Spark 1.1 Pricing: What Heavy AI Users Need to Know

Meta just did something it has never done before: charge for a model.

Muse Spark 1.1 launched on July 9, 2026, and with it came the first-ever paid Meta Model API. The pricing is $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits to get started. That single number changes the competitive landscape for every heavy AI user running code-heavy or agentic workflows in 2026.

Here is what you need to know before you route a single token its way.

Muse Spark Pricing: $1.25 In, $4.25 Out

The headline rate sits comfortably below Claude Sonnet 5’s intro pricing ($2/$10 per MTok) and well below Claude Opus 4.8 ($5/$25 per MTok). It is roughly on par with Grok 4.5, which Cursor launched the same day at $2/$6 per MTok.

Quick reference:

Model	Input $/MTok	Output $/MTok
Claude Opus 4.8	$5.00	$25.00
GPT-5.5	$3.00	$15.00
Claude Sonnet 5 (intro)	$2.00	$10.00
Grok 4.5	$2.00	$6.00
Muse Spark 1.1	$1.25	$4.25
DeepSeek V4 Pro	$0.43	$1.10

For a workload generating 1 million output tokens per day, that is a $4.25 daily bill versus $10 on Sonnet 5 intro and $25 on Opus 4.8. Monthly, you are looking at roughly $127 versus $300 versus $750. The gap compounds fast at scale.

The 262k context window matches what you get on Claude Sonnet 5. So for long-context agentic pipelines, Muse Spark 1.1 is not giving anything up on that axis.

What Muse Spark 1.1 Actually Is

Muse Spark 1.1 is the API-ready version of the Muse Spark model Meta first released in April 2026. The original Muse Spark was hosted-only and free: useful for testing, irrelevant for production. Version 1.1 changes that entirely.

Meta claims significant improvements in agentic tool calling and computer use over the April release. The model supports text, image, and speech input, outputs text, and operates in both a standard and reasoning mode. It has a 262k token context window and exposes a standard chat completions-style API.

Simon Willison, who had preview access, noted that the model shows “major gains in agentic tool calling and computer use” and that the Muse Spark 1.1 Evaluation Report has more depth on what changed. The report apparently includes a section on “Attractor States in Self-Conversation” where two copies of the model talking to each other produce notably consistent philosophical loops — an interesting stress test of internal coherence.

The April version benchmarked competitively with Claude Opus 4.6 on selected tasks but showed performance gaps in coding and long-horizon agentic workflows. Meta explicitly acknowledged those gaps in April and said they were working on them. Muse Spark 1.1 is the answer, and the timing — same week Cursor launched Grok 4.5 — signals that Meta is treating the developer API market as a direct battleground.

The SemiAnalysis Critique You Should Know

One piece of context that matters for heavy users evaluating Muse Spark 1.1: SemiAnalysis published a piece this week noting that despite Meta running five Titan compute clusters above 1GW and deploying roughly 3,000 engineers on reinforcement learning environments, April’s Muse Spark “still only performs at Opus 4.6 level.”

That is a pointed critique. Opus 4.6 is two versions behind Claude’s current flagship. If Muse Spark 1.1 has only modestly improved on that baseline, then the pricing advantage is real but the capability ceiling matters for tasks where you need frontier performance.

For coding tasks specifically, independent benchmarks have not yet caught up to the 1.1 release. The April version was behind GPT-5.5 and Claude on SWE-Bench style evaluations. If 1.1 closes that gap materially, the value proposition sharpens considerably. If it does not, then Muse Spark 1.1 is best positioned as a high-throughput, cost-sensitive layer rather than a primary reasoning engine.

Where Muse Spark 1.1 Makes Sense in a Multi-Provider Stack

For users already running a tiered routing strategy, Muse Spark 1.1 opens a slot between DeepSeek V4 Pro ($0.43/$1.10) and Claude Sonnet 5 intro ($2/$10). That middle tier is useful for:

Classification and routing tasks. If you are using a model to decide which downstream model handles a user request, you want speed and low cost, not Opus-class reasoning. Muse Spark 1.1 at $1.25/$4.25 is plausible here, though you would want to run your own accuracy benchmarks before committing.

High-volume summarization. Summarizing meeting transcripts, documents, or codebases at scale is output-heavy. At $4.25 per million output tokens versus $10 on Sonnet 5, that is a 2.4x cost reduction per summary unit, assuming comparable quality.

Agentic scaffolding calls. In multi-step agent loops, a large share of your token spend goes to context management, intermediate reasoning, and tool call parsing. These are tasks where reasoning quality matters less than throughput and reliability. Muse Spark 1.1 could absorb a meaningful fraction of those calls.

The caveat: Meta’s API infrastructure is brand new. $20 in free credits is a generous sandbox, but production reliability and rate limits are unknown quantities at launch. Early adopters should expect the usual first-month friction: undocumented rate limits, occasional latency spikes, and the inevitable round of SDKs catching up to the new endpoint.

Free Credits and the Acquisition Play

The $20 free credit offer is not accidental. It mirrors the strategy OpenAI and Anthropic used in their early API growth phases: lower the barrier to first production call, let usage patterns develop, and then convert on volume.

For heavy users already spending $300 to $1,000 per month on AI APIs, $20 is about two hours of production load at medium intensity. It is enough to run a meaningful evaluation but not enough to replace a real commitment. The pricing needs to hold up at scale, and that means independent benchmarking over the next few weeks is going to determine whether Muse Spark 1.1 is a primary stack component or a specialty tool.

Cost Math for a Typical Heavy User

Assume you are running a coding assistant workload: roughly 2 million input tokens and 500,000 output tokens per day, five days a week.

On Claude Sonnet 5 intro pricing: (2 x $2.00) + (0.5 x $10.00) = $9.00 per day, $45 per week, $180 per month.

On Muse Spark 1.1: (2 x $1.25) + (0.5 x $4.25) = $4.625 per day, $23.13 per week, $92.50 per month.

That is roughly $87.50 saved per month on that workload alone, assuming equivalent output quality. If you run multiple workloads or higher volumes, the number scales linearly.

What to Watch in the Next Two Weeks

Meta’s entry into the commercial API market is the first meaningful structural change to the pricing landscape since DeepSeek V4 Pro landed in June. Three things will determine whether Muse Spark 1.1 becomes a genuine Claude and GPT competitor for heavy users:

1. Independent coding benchmarks. SWE-Bench Pro, Cursor bench, and LiveCodeBench results for 1.1 need to appear. Meta’s self-reported numbers are a starting point, not a verdict.

2. Rate limit transparency. The initial API preview is developer-facing. What limits apply at $1.25/$4.25 pricing, and what happens when you hit them, will define whether this is production-grade.

3. SDK and integration maturity. Simon Willison already released an LLM plugin for Muse Spark 1.1 within hours of launch. Vercel, LangChain, and other orchestration layers will follow. How fast that ecosystem matures matters for teams that do not want to build raw HTTP integrations.

The short version: Muse Spark 1.1 pricing is genuinely competitive. Meta is serious about the commercial API market. But production readiness for heavy workloads needs four to six weeks of community testing before the risk profile is clear enough to commit budget.

If you are tracking your per-provider token spend across Claude, GPT, and Grok, this is the week to add a Muse Spark column to your cost dashboard.