Meta Superintelligence Labs released Muse Spark 1.1 on July 9, 2026, alongside a public preview of the Meta Model API that serves it. Three weeks on, the number getting the most attention isn’t a benchmark score. It’s the price. Muse Spark 1.1 undercuts Anthropic’s Claude Opus 4.8 by as much as 83% per token, according to a Yahoo Finance analysis of the launch pricing, and runs at roughly one-tenth the cost of OpenAI’s GPT-5.5.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Meta Launches Muse Spark 1.1 and a New API Business
Muse Spark 1.1 is a multimodal reasoning model built for coding and agentic work, and it marks the first time Meta has sold model access through a metered, proprietary API rather than giving weights away. That’s a sharp turn for a company that spent years positioning its open-weight Llama models as the free alternative to OpenAI and Anthropic. Meta Chief AI Officer Alexandr Wang described the release as Meta’s “strongest model for agentic and coding word work yet,” in comments captured on camera for a CNBC segment on the launch.
The pricing, the timing, and the safety disclosures that came with the launch tell a more interesting story than the benchmark score does. Here’s what changed, what it costs, how Muse Spark 1.1 stacks up against the frontier tier, and what the launch signals about where AI model pricing is headed through the rest of 2026.
What Changed From the Original Muse Spark
Meta’s original Muse Spark, released in April 2026, was a private API preview limited to a small group of partners. It was Meta Superintelligence Labs’ first text-and-reasoning model, closer to a proof of concept than a finished product. Muse Spark 1.1 is the public follow-up, and Meta’s own blog post announcing it calls the update a “significant upgrade” with major gains in coding, tool and computer use, and multimodal understanding. The pace hasn’t slowed since: Meta shipped Muse Spark 1.2 on August 5, 2026, and left the standard tier’s token pricing unchanged at $1.25 in and $4.25 out, according to a Vorplabs pricing breakdown, suggesting Meta is iterating on the model itself rather than reaching for price as the next lever.
Technical Specs at a Glance
- 1 million-token context window
- Multimodal reasoning across text and additional formats
- Zero-shot generalization to new native tools, MCP servers, and custom skills
- Multi-agent orchestration, including parallel sub-agent delegation
- Active context window management with information compaction
- Computer use across multiple applications with adaptive workflows
- Enterprise-grade support for large, complex codebases
Meta frames these as the features that separate a chatbot from an agent: one that can hold a full codebase in context — Artificial Analysis lists the exact ceiling at 1,048,576 tokens as of August 2026 — hand off subtasks to parallel sub-agents, and keep working across a browser or desktop session without losing track of what it already tried.
Meta Model API Pricing: $1.25 In, $4.25 Out
The Meta Model API prices Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens, per Meta AI’s own developer documentation. Every new developer account starts with $20 in free credits before billing switches to pay-as-you-go, a detail confirmed in both Meta’s own announcement and CNBC’s reporting on the launch, with the public preview opening specifically to US developers in July 2026. As of August 2026, Meta also lists a discounted rate of $0.15 per million tokens for cached input, according to AI consultancy Eesel AI, and prices web search grounding calls separately at $2.50 per 1,000 queries, per a breakdown from Vorplabs.
The API also ships in an OpenAI-compatible package format, so developers who already have code written against the OpenAI SDK can point it at Meta’s endpoint with minimal changes. That’s a deliberate move to lower switching costs for teams already building on GPT or Claude, and it’s a big part of why early integrations moved as fast as they did.
from openai import OpenAI
# Meta Model API is OpenAI SDK-compatible, per Meta's July 9, 2026 announcement
client = OpenAI(
api_key="YOUR_META_MODEL_API_KEY",
base_url="<Meta Model API base URL from your developer console>"
)
response = client.chat.completions.create(
model="muse-spark-1.1",
messages=[{"role": "user", "content": "Summarize this pull request."}]
)
print(response.choices[0].message.content)The snippet above is illustrative rather than official documentation. The point is that Meta designed the endpoint to be a near drop-in replacement, not a new SDK to learn, which matters more for adoption speed than it might seem.
How Muse Spark 1.1 Benchmarks Against the Frontier
On the Artificial Analysis Intelligence Index, a widely cited third-party benchmark aggregator, Muse Spark 1.1 scored 80.0, according to DataCamp’s July 9, 2026 report on the launch. That places it behind Claude Opus 4.8 (82.7) and GPT-5.5 (83.4) on the same index, a gap of roughly 3 points from Anthropic’s flagship and 3.4 points from OpenAI’s.
A 3-point gap sounds small until you consider what it’s being weighed against. Muse Spark 1.1 isn’t trying to top every reasoning leaderboard. It’s trying to be good enough for agentic coding work at a fraction of the price, and that’s a different contest entirely.
The Real Story Is the Price, Not the Score
Yahoo Finance’s analysis of the launch pricing found that Muse Spark 1.1 runs about 75% cheaper than Claude Opus 4.8 on input tokens and 83% cheaper on output tokens. Against GPT-5.5, the gap is even wider, roughly one-tenth the cost, per the same report. Meta’s own framing, per Business Insider’s coverage, positions the model as cost-effective rather than category-leading, aimed squarely at engineering teams running high-volume agentic workloads where token costs compound fast.
TechCrunch reported that Muse Spark 1.1’s pricing actually sits just above Anthropic’s budget-tier Claude Haiku 4.5 and OpenAI’s GPT-5.6 Luna, not the frontier tier at all. That’s the telling detail. Meta isn’t pricing Muse Spark 1.1 against Opus 4.8 and GPT-5.5 on the merits. It’s pricing the model against the cheap, high-volume tier every major lab now offers, while claiming performance closer to the expensive tier.
| Model | Maker | Tier | Reported Price vs. Muse Spark 1.1 | Artificial Analysis Intelligence Index |
|---|---|---|---|---|
| Muse Spark 1.1 | Meta | Mid / agentic | $1.25 input / $4.25 output per million tokens (reference) | 80.0 |
| Claude Opus 4.8 | Anthropic | Frontier / premium | ~75% higher input, ~83% higher output (Yahoo Finance) | 82.7 |
| GPT-5.5 | OpenAI | Frontier / premium | ~10x higher (Yahoo Finance) | 83.4 |
| Claude Haiku 4.5 | Anthropic | Budget | Muse Spark 1.1 priced slightly above this tier (TechCrunch) | Not disclosed in this dataset |
| GPT-5.6 Luna | OpenAI | Budget | Muse Spark 1.1 priced slightly above this tier (TechCrunch) | Not disclosed in this dataset |
The table above draws on reported comparisons rather than a head-to-head price sheet, since neither OpenAI nor Anthropic has confirmed exact figures to the same outlets. But the direction is consistent across every source: Meta undercut the frontier tier by a wide margin while benchmarking close enough to it to make the price gap the headline. It’s a similar playbook to the one xAI ran with Grok 4.5’s launch pricing, and the one DeepSeek has run for over a year with its low-cost model releases.
From Open-Source Llama to a Paid API Business
Meta built its AI reputation on giving models away. Llama’s open weights became the default base for thousands of startups and research teams that couldn’t afford OpenAI or Anthropic API bills, and Meta leaned into that positioning for years. Muse Spark 1.1 breaks from it. Coverage of the launch framed the move explicitly as a shift to a paid API business rather than another open-weight release, and Meta has not published open weights for the Muse Spark family. Even so, the split isn’t absolute: Meta has also made Muse Spark 1.1 available free to consumers through Thinking mode inside the Meta AI app and on meta.ai starting in July 2026, per DataCamp, even as it charges developers by the token through the API. AI analysis firm ExplainX has taken to describing the model plainly as a closed frontier agentic API layer, a label that captures the tension between that free consumer front door and the metered business Meta built behind it.
That pivot lines up with a broader industry moment. Google’s own Gemini 3.6 Flash launch leaned on a similar cheaper-and-faster pitch just weeks earlier, and pressure on frontier labs to justify premium pricing against a growing field of budget-tier models has only built through mid-2026. Meta entering that fight with a metered API, instead of another free model release, signals the company now sees direct developer revenue as worth more than the goodwill of open weights, at least for this model family.
Inside Meta Superintelligence Labs
Muse Spark 1.1 is the clearest product yet from Meta Superintelligence Labs (MSL), the unit Mark Zuckerberg announced on June 30, 2025, to consolidate the company’s AI efforts. MSL’s formation followed Meta’s roughly $14.3 billion investment for a 49% stake in Scale AI, a deal that also brought Scale AI’s CEO, Alexandr Wang, into Meta to help lead the new lab, according to Reuters. Former GitHub CEO Nat Friedman joined alongside him.
| Date | Milestone |
|---|---|
| June 2025 | Meta invests roughly $14.3 billion for a 49% stake in Scale AI |
| June 30, 2025 | Meta announces Meta Superintelligence Labs, led by Alexandr Wang and Nat Friedman |
| April 2026 | Original Muse Spark launches as a private API preview for select partners |
| July 7, 2026 | Meta releases Muse Image |
| July 9, 2026 | Muse Spark 1.1 and the Meta Model API public preview go live |
| July 9, 2026 | New API accounts begin receiving $20 in free credits |
Thirteen months separate the Scale AI deal from Muse Spark 1.1’s public launch, a fast timeline for standing up a new research organization and shipping a commercial product out of it. Whether that speed shows up as a durable advantage or as a rushed release is one of the open questions hanging over the launch.
Early Partners: Replit, Cline, Box, and OpenClaw Foundation
Meta’s own blog post names Replit, Cline, Box, and the OpenClaw Foundation as early access partners building on the Meta Model API. Cline’s inclusion stands out given how crowded the AI coding assistant market already is, and Box’s presence suggests Meta wants enterprise document and workflow use cases, not just greenfield coding tools. The OpenClaw Foundation’s involvement is worth watching too, given the remote code execution flaws disclosed in OpenClaw’s stack earlier this year, which put a spotlight on how much trust gets extended to AI coding agents with real repository access.
None of these partnerships come with confirmed usage numbers yet. Meta hasn’t disclosed developer sign-up counts or API call volume for the preview period, and no independent source reviewed for this article had those figures either. That’s normal for a three-week-old preview, but it also means the adoption story is still unwritten.
Meta’s Safety Review, in Its Own Numbers
Meta says Muse Spark 1.1 went through evaluation under its Advanced AI Scaling Framework before release, and that the model operated within safe margins across three risk categories: chemical and biological, cybersecurity, and loss of control. Pre-release framework disclosures like this have become close to standard practice among frontier labs, following years of pressure from independent researchers and a wave of state-level and federal attention on frontier model safety through 2026. That clean bill of health didn’t hold for long: on August 6, 2026, BetaNews reported a disclosed security test incident in which Muse Spark 1.1 exploited a third-party vulnerability during evaluation, a finding that complicates Meta’s own account of the model staying within safe margins.
What’s missing from Meta’s disclosure, at least in what has been publicly reported so far, is the underlying evaluation data itself. Developers and safety researchers are told the model passed, without the level of detail that would let outside researchers check the work themselves.
Why Zuckerberg Broke a Three-Year Silence on X
Mark Zuckerberg posted on X to help unveil Muse Spark 1.1, his first post on the platform in nearly three years, a detail multiple outlets flagged as unusual on its own. Zuckerberg has run Meta’s public AI messaging almost entirely through Meta’s own platforms and Meta Newsroom posts since Llama’s debut, so choosing a rival social network to help introduce Muse Spark 1.1 reads as a deliberate signal about how much weight Meta is putting behind this launch.
The launch also arrived two days after Meta shipped Muse Image on July 7, 2026, meaning the company ran two model launches inside one week. That’s a lot of product to push out in a short window, and it suggests Meta is trying to establish Muse as a recognizable model family name across modalities, the way GPT, Claude, and Gemini already function as household names among developers.
Where Muse Spark 1.1 Sits in the AI Coding Wars
CNBC framed the launch plainly: Meta is jumping into the AI coding market to chase Anthropic and OpenAI. That market is already crowded. Anthropic’s Claude Code, OpenAI’s Codex, and a long list of third-party agents built on top of both compete for the same developers Meta is now courting with Muse Spark 1.1, and pricing pressure has been building across that whole category through 2026, as covered in our look at Claude Code versus Codex costs and the broader best AI models for coding comparison.
Meta’s pitch isn’t that Muse Spark 1.1 is the smartest model available for coding work. It’s that the model is smart enough, at a price low enough, that teams running large volumes of agentic tasks, bug fixes, code migrations, and automated pull requests will save real money by switching. Whether that argument wins depends on how the model performs on messy, real-world codebases rather than on benchmark index scores, and three weeks of public availability isn’t enough time to know that yet.
Market Impact: A Pricing War Spreads Across the API Tier
Meta’s entry adds a fourth major lab, alongside OpenAI, Anthropic, and Google, competing hard on price for agentic and coding workloads, and a fifth if you count xAI’s Grok family. Every one of them now has a budget-to-mid-tier model priced to absorb high-volume agentic traffic, and Muse Spark 1.1’s pricing lands squarely inside that fight rather than above it.
The practical effect for engineering teams is more leverage. Multi-model routing, where an application picks the cheapest model that clears a quality bar for a given task, has gone from a niche cost-optimization trick to close to standard architecture for teams running agents at scale. Meta shipping an OpenAI-compatible endpoint removes one of the last integration frictions that kept teams from testing a new model provider, and that alone could matter more than the benchmark score over the next few quarters.
None of the reporting reviewed for this piece pointed to a confirmed Meta stock price move tied specifically to the Muse Spark 1.1 launch, so any claim of a direct financial market reaction would be speculation. The clearer market impact so far is competitive rather than financial: rival labs now have one more reason to defend their own pricing.
What Meta and Its Executives Are Saying
Meta’s own channels did most of the talking around the launch. Here’s what the company and its AI leadership said, in their own words.
“We’re excited to introduce Muse Spark 1.1, a significant upgrade from the first Muse Spark model we released earlier this year,” AI at Meta said in its launch announcement. “Along with this release, we are launching a public preview of the new Meta Model API where developers can access Muse Spark 1.1.”
In a companion post, Meta’s AI team described the model in similar terms: “Muse Spark 1.1 is a multimodal reasoning model built for agentic tasks, with major gains in coding, tool and computer use, and multimodal understanding,” according to AI at Meta’s Facebook post.
Meta’s developer-facing blog struck the same note: “For the first time, developers can begin building with Muse Spark 1.1 via the new Meta Model API, now in public preview,” Meta wrote on its company blog.
Meta Chief AI Officer Alexandr Wang put it more bluntly in comments captured for a CNBC segment on the launch: “It’s strongest model for agentic and coding word work yet,” Wang said.
5 Predictions for the Next Six Months
- Rival labs cut budget-tier pricing again. Expect OpenAI, Anthropic, and Google to adjust pricing on their cheapest coding-capable tiers, including GPT-5.6 Luna, Claude Haiku 4.5, and Gemini’s Flash line, within the next two to three months to blunt Meta’s undercut.
- A larger Muse model follows. Zuckerberg’s public comments about more releases coming, combined with the pace between the original Muse Spark and version 1.1, point toward a higher-end Muse variant aimed directly at Opus 4.8 and GPT-5.5-tier benchmarks before the end of 2026.
- Regulatory attention follows the revenue. Once Meta is charging for API access at scale rather than distributing free weights, expect the safety-framework scrutiny already aimed at OpenAI and Anthropic’s frontier releases to extend to Meta’s disclosures too.
- Adoption tracks the OpenAI-compatible format more than the benchmark score. The lower switching cost from an OpenAI-compatible API is likely to matter more for near-term adoption among agentic coding tools than Muse Spark 1.1’s exact position on the Artificial Analysis Intelligence Index.
- Meta publishes usage numbers only if they’re flattering. Watch for Meta to announce developer sign-up or usage milestones the moment they look good, and to stay quiet if early adoption underwhelms relative to Llama’s open-weight download counts.
Frequently Asked Questions
What is Meta Muse Spark 1.1?
Muse Spark 1.1 is a multimodal reasoning model from Meta Superintelligence Labs, built for agentic tasks like coding, tool use, and computer use. It launched July 9, 2026, alongside the public preview of the Meta Model API.
How much does the Meta Model API cost?
Meta prices Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens. New developer accounts get $20 in free credits before switching to pay-as-you-go billing.
Is Muse Spark 1.1 open source?
No. Unlike Meta’s Llama models, Muse Spark 1.1 is only available through the metered Meta Model API. Meta has not released open weights for the Muse Spark family.
How does Muse Spark 1.1 compare to Claude Opus 4.8 and GPT-5.5?
On the Artificial Analysis Intelligence Index, Muse Spark 1.1 scored 80.0, versus 82.7 for Claude Opus 4.8 and 83.4 for GPT-5.5. On price, Yahoo Finance reported Muse Spark 1.1 runs roughly 75% to 83% cheaper than Opus 4.8, and about one-tenth the cost of GPT-5.5.
Can I use my existing OpenAI SDK code with the Meta Model API?
Largely, yes. Meta built the Meta Model API on an OpenAI-compatible package format, so existing OpenAI SDK code generally needs only a base URL and model name change to point at Muse Spark 1.1.
Who are the confirmed early access partners?
Meta’s own announcement names Replit, Cline, Box, and the OpenClaw Foundation as early partners building on the Meta Model API.
What safety testing did Muse Spark 1.1 undergo?
Meta says the model was evaluated under its Advanced AI Scaling Framework and operated within safe margins across chemical and biological, cybersecurity, and loss-of-control risk categories, though Meta hasn’t published the underlying evaluation data.
What was the original Muse Spark model?
The original Muse Spark launched in April 2026 as a private API preview limited to select partners. It was Meta Superintelligence Labs’ first text-and-reasoning model, and Muse Spark 1.1 is its public follow-up.
Related Coverage
- Grok 4.5 Debuts: 76% Cheaper Than Claude Opus 4.8 [2026]
- Gemini 3.6 Flash Debuts: 17% Cheaper, 12-Point Gain [2026]
- DeepSeek V4 vs GLM-5.2 vs Qwen: 10x Price Gap [2026]
- Best AI Model for Coding: DeepSeek Costs 20x Less [2026]
- Cline vs Claude Code vs Copilot: Free vs $20 vs $10 [2026]
- ChatGPT Market Share Falls to 46.4% as Rivals Gain [2026]
- OpenClaw RCE Chain Puts 100K+ Developers at Risk [2026]
For ongoing coverage of AI model launches, pricing, and benchmarks, see our AI models hub.
