Meta Muse Spark 1.1: A Cheap Agentic Coding Bet
Meta launched Muse Spark 1.1 at 1.25 and 4.25 dollars per million tokens to chase Anthropic and OpenAI. Pricing, benchmarks, and the honest verdict.
By Capital & Compute
Meta shipped Muse Spark 1.1 on July 9, 2026, and did something it had never done before: it charged for it. Meta’s first paid model arrives at $1.25 per million input tokens and $4.25 per million output, an aggressive wedge aimed straight at the API market Anthropic and OpenAI built. The launch benchmarks tell a sharper story than the price does. Muse Spark leads every tool-use and agentic test Meta published and trails the frontier on every pure-coding one. It is the cheapest capable agent-runner on the board, and it is not the best coder. Those are different claims, and Meta is betting the first one matters more.
| Metric | Muse Spark 1.1 | Claude Opus 4.8 |
|---|---|---|
| MCP Atlas (scaled tool use) | 88.1 | 82.2 |
| JobBench (pro tool use) | 54.7 | 48.4 |
| Humanity's Last Exam (tools) | 62.1 | 57.9 |
| OSWorld-Verified (computer use) | 80.8 | 83.4 |
| SWE-Bench Pro (coding) | 61.5 | 69.2 |
| DeepSWE 1.1 (long-horizon) | 53.3 | 59.0 |
| BabyVision (visual reasoning) | 76.3 | 81.2 |
What Meta actually shipped
Muse Spark 1.1 is a multimodal reasoning model built for agentic tasks, released by Meta Superintelligence Labs and announced on Meta’s developer blog (Introducing Muse Spark 1.1 and the Meta Model API, July 9, 2026). It carries a 1 million-token context window and is pitched at tool use, computer use, coding, and multimodal understanding. Meta says it “zero-shot generalizes to new native tools, MCP servers, and custom skills” and can orchestrate multi-agent systems. The model runs free in the “Thinking” mode of the Meta AI app and at meta.ai, while the paid Meta Model API opened as a US-only public preview.
The context that makes this a market event, not just a model drop: it is Meta’s first model that costs money to call. As CNBC reported, this is Meta chasing Anthropic and OpenAI three months after AI chief Alexandr Wang unveiled the first Muse Spark, and Wang called 1.1 the company’s strongest model for agentic and coding work yet. Mark Zuckerberg framed the pitch in one line, per TechCrunch: “a strong agentic and coding model at a very low price.”
Meta also lowered the cost of trying it. New API accounts start with $20 in free credits, and the Model API is designed to accept both OpenAI and Anthropic request formats, as detailed in developer coverage of the launch: point the base URL at api.meta.ai/v1, pass a key, name the model muse-spark-1.1. That dual compatibility is a switching-cost play. Code already written against a competitor SDK can call Meta with a one-line change, which is exactly the friction a late entrant needs to remove.
The price wedge
The whole strategy is in the rate card. Here is where Muse Spark lands against the models it is priced against and the frontier it is chasing, per-million-token list rates.
| Model (maker) | Input / Mtok | Output / Mtok |
|---|---|---|
| GPT-5.6 Luna (OpenAI) | $1.00 | $6.00 |
| Claude Haiku 4.5 (Anthropic) | $1.00 | $5.00 |
| Muse Spark 1.1 (Meta) | $1.25 | $4.25 |
| Claude Opus 4.8 (Anthropic) | $5.00 | $25.00 |
| Claude Fable 5 (Anthropic) | $10.00 | $50.00 |
CNBC described Muse Spark as priced “in line with, albeit slightly above” GPT-5.6 Luna and Claude Haiku 4.5. That is true on the input sticker and misleading on the bill. Muse Spark’s input is a quarter above the $1.00 that Luna and Haiku charge, but its output rate of $4.25 comes in below both Luna’s $6.00 and Haiku’s $5.00. Agentic and coding work is output-heavy: the model spends most of its tokens generating plans, tool calls, and code, not reading the prompt. On a typical input-heavy-then-output-heavy agent turn, Muse Spark is at or below the cost of both budget rivals, not above them.
Against the frontier the gap is not subtle. Muse Spark’s output rate is roughly a sixth of Claude Opus 4.8 at $25 and less than a tenth of Claude Fable 5 at $50. This is the same script that SpaceXAI ran with Grok 4.5 a day earlier and the one that has made Chinese open-weight models so cheap: get close enough on capability to be credible, then make the comparison about price. The difference is that Meta is not a scrappy challenger. It is spending frontier-lab money to enter as the low-cost option on purpose.
What the benchmarks really say
Meta published a scorecard with the launch. The numbers below are Meta’s own, drawn from its launch charts and compiled by MarkTechPost; the (max) and (xhigh) settings are the reasoning-effort levels Meta reported for each model.
| Benchmark (what it tests) | Muse Spark 1.1 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| MCP Atlas (scaled tool use) | 88.1 | 82.2 | 75.3 | 78.2 |
| JobBench (professional tool use) | 54.7 | 48.4 | 38.3 | 15.9 |
| Humanity’s Last Exam (reasoning + tools) | 62.1 | 57.9 | 52.2 | 51.4 |
| OSWorld-Verified (computer use) | 80.8 | 83.4 | 78.7 | 76.2 |
| SWE-Bench Pro (real-repo coding) | 61.5 | 69.2 | 58.6 | 54.2 |
| DeepSWE 1.1 (long-horizon coding) | 53.3 | 59.0 | 67.0 | 12.0 |
| BabyVision (visual reasoning) | 76.3 | 81.2 | 83.6 | 51.5 |
The pattern is clean enough to be a product strategy. Muse Spark tops the three benchmarks about calling tools and running agents. It leads the field on MCP Atlas, a scaled tool-use test, and it wins JobBench, a professional tool-use suite, by more than six points over Opus 4.8 and by nearly 39 over Gemini 3.1 Pro. On Humanity’s Last Exam run with tools, it leads again.
Then the picture inverts on the work that actually ships software. On SWE-Bench Pro, which grades resolving real issues inside a large existing repository, Muse Spark scores 61.5 against Opus 4.8 at 69.2, an eight-point deficit. On DeepSWE 1.1, a long-horizon agentic coding benchmark, it sits at 53.3 while GPT-5.5 leads at 67.0 and Opus 4.8 reaches 59.0. It trails on computer use (OSWorld-Verified) and on visual reasoning (BabyVision) too. So the model that is best at deciding which tool to call is not the best at writing the code that tool runs.
The missing ecosystem
A model is not a product. The tools developers actually reach for, Claude Code and the rest of the agent field, are harnesses wrapped around a model: a CLI or desktop app, IDE integrations, memory, permissions, and customization. Beating them on a coding benchmark is not the same as replacing them, and Muse Spark launched as an API and a chat mode, with no first-party coding agent, desktop app, or editor integration announced.
What Meta shipped instead is standards support and partners. The MCP Atlas lead is not incidental: Meta is betting on the Model Context Protocol as the connective tissue, so any MCP-native harness can point at Muse Spark without Meta having to build the harness itself. The launch came with endorsements from three companies that build those harnesses, quoted on Meta’s blog: Replit’s Amjad Masad, Cline’s Saoud Rizwan, and Box’s Yashodha Bhavnani. That is the adoption path. Meta supplies a cheap, tool-fluent engine, the ecosystem supplies the cockpit, and the OpenAI and Anthropic SDK compatibility lets that ecosystem wire it in without a rewrite.
It is a coherent bet, and it has a hole. The developer who lives inside Claude Code today gets the model and the harness from one vendor, tightly integrated. To adopt Muse Spark, that developer needs a third-party agent that supports it well, and on launch day the strongest such integrations are still being built. Cheap and tool-fluent gets Meta into the conversation. It does not yet get it into the terminal.
The bottom line
Muse Spark 1.1 is the clearest signal yet that the AI coding market is now a price war fought on a second front. The first front, raw coding accuracy, still belongs to Anthropic and OpenAI, and Meta’s own benchmarks concede it. The second front is agentic orchestration at volume: how cheaply a model can run an agent that calls tools, chains steps, and works across applications. On that front Muse Spark leads and it is priced to win, with an output rate below the budget tier and a fraction of the frontier.
For anyone budgeting AI work, the honest read is narrow and useful. If the job is a high-volume agent that spends its tokens calling tools and managing workflows, Muse Spark is worth a real trial, and the $20 in free credits makes that trial free. If the job is resolving hard issues inside a large codebase, the frontier coders still finish more of them, and the cost per finished task, not the per-token rate, is the number to run before switching. Meta has bought its way to the table. Whether it earns a seat depends on whether the harnesses show up, and on whether tool-use skill turns out to matter as much as Meta is wagering it does. Watch the AI model release tracker for the follow-on releases Zuckerberg promised, and the coding agents hub for the integrations that would turn this model into a tool people actually code in.
Frequently asked questions
- How much does Meta Muse Spark 1.1 cost?
- The Meta Model API lists Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits for new accounts. That output rate is below GPT-5.6 Luna ($6) and Claude Haiku 4.5 ($5), and roughly a sixth of Claude Opus 4.8 ($25). It is Meta's first paid model.
- Is Muse Spark 1.1 good for coding?
- It is capable but not best-in-class. On Meta's own launch benchmarks it scores 61.5 on SWE-Bench Pro and 53.3 on DeepSWE 1.1, trailing Claude Opus 4.8 (69.2 and 59.0) and GPT-5.5 on long-horizon coding. Where it leads is tool use and agentic orchestration, not pure code accuracy.
- How does Muse Spark compare to Claude and GPT?
- Muse Spark 1.1 tops Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on the three tool-use and agentic benchmarks Meta published (MCP Atlas, JobBench, Humanity's Last Exam with tools) and trails Opus 4.8 on real-repo coding, computer use, and visual reasoning. It is the cheapest of the group per output token.
- Can I use Muse Spark 1.1 in an IDE like Cursor or Claude Code?
- Not as a first-party product at launch. Meta shipped an API and a chat mode, with no desktop app or IDE integration of its own. Adoption depends on third-party harnesses, and the API accepts both OpenAI and Anthropic request formats to make that wiring easier.
- Why did Meta launch a paid coding model?
- To plant a flag in the frontier-model API market that Anthropic and OpenAI dominate. Charging for Muse Spark 1.1 is Meta entering that market directly, and pricing it as the low-cost, tool-fluent option is the wedge it chose to compete on.
Sources
Meta Superintelligence Labs (2026). Introducing Muse Spark 1.1 and the Meta Model API. Meta AI developer blog. https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/
CNBC (2026, July 9). Meta jumps into AI coding market in effort to chase Anthropic and OpenAI. CNBC. https://www.cnbc.com/2026/07/09/meta-jumps-into-ai-coding-market-to-chase-anthropic-and-openai.html
TechCrunch (2026, July 9). Meta enters the crowded AI coding battle with Muse Spark 1.1. TechCrunch. https://techcrunch.com/2026/07/09/meta-enters-the-crowded-ai-coding-battle-with-muse-spark-1-1/
MarkTechPost (2026, July 9). Meta Superintelligence Labs Releases Muse Spark 1.1: A Multimodal Reasoning Model for Agentic Tasks on Meta Model API. MarkTechPost. https://www.marktechpost.com/2026/07/09/meta-superintelligence-labs-releases-muse-spark-1-1/
Digital Applied (2026). Meta Muse Spark 1.1: Meta’s First Paid Agent Model. Digital Applied. https://www.digitalapplied.com/blog/meta-muse-spark-1-1-agentic-model-api-2026