Meta’s Muse Spark 1.1 Beats Rivals on Price, Trails on Coding

Meta opened its Muse Spark 1.1 coding model to developers in the United States this week, pricing it at $1.25 per million input tokens and $4.25 per million output tokens. That undercuts the going rates at Anthropic and OpenAI by a wide margin, and it is the first time Meta has ever charged outside developers to use one of its own AI models.

The discount shrinks once the model actually starts thinking. Muse Spark 1.1 bills its own chain of reasoning at that same $4.25 output rate, and Meta’s own benchmark table already shows the model trailing Claude Opus 4.8 and GPT-5.5 on the coding tests developers watch most closely. European developers cannot test any of it yet.

A Coding Model Priced Below Anthropic and OpenAI

Muse Spark 1.1 writes and debugs code, calls outside software tools, reads text, images, video and documents, and carries out multi-step jobs with little human oversight, according to Meta. It also generalizes to new tools, Model Context Protocol (MCP) servers and custom skills without extra training, and it carries a context window of one million tokens, large enough to hold several codebases in memory at once.

Every new account on Meta’s developer platform starts with $20 in free credits before metered pricing kicks in. The company launched a public preview of the Meta Model API on Thursday, giving outside developers their first hosted access to a Meta frontier model. From there, it costs $1.25 per million input tokens and $4.25 per million output tokens.

Reuters reported the rate sits above OpenAI’s entry level GPT-5 mini and Anthropic’s low cost Claude Haiku 4.5, but below Anthropic’s higher end Claude Sonnet 4.6. Anthropic’s own introductory rate for Claude Sonnet 5 runs $2 per million input tokens and $10 per million output tokens through August 31, 2026, before rising to $3 and $15. Claude Opus 4.8 costs $5 per million input tokens and $25 per million output tokens. OpenAI has not announced broad pricing for its GPT-5.6 model.

Mark Zuckerberg, Meta’s chief executive, called Muse Spark 1.1 “a strong agentic and coding model at a very low price” in a post on X, adding it was “strongest at agentic performance, tool use, and computer use.” Alexandr Wang, Meta’s chief AI officer, told CNBC the pricing is “very aggressive and attractive” and said the goal is “to really have attractive pricing that scales with immense consumption usage.”

Zuckerberg has framed the strategy as a rebuttal to how rivals price their own models. “The pricing from some of the other labs is very extreme and has very high margins,” he said. “We think that there’s a real ability to be able to offer frontier or very high level intelligence at a much more affordable cost.”

The Reasoning Tokens That Erase the Discount

Muse Spark 1.1 is a reasoning model, which means it generates internal chain of thought tokens before it produces a visible answer. Meta bills those thinking tokens at the full $4.25 output rate, not a discounted one.

Developers can dial reasoning depth up or down with an effort parameter, the main lever for controlling cost on harder tasks. The headline price holds for simple agentic calls. For compute heavy reasoning work, the effective cost per finished task climbs well past the sticker figure.

Cached input tokens are billed separately at $0.15 per million, a steep discount aimed at agent loops that repeat the same context turn after turn. One developer cost breakdown estimated a busy multi agent session generating 4 million output tokens in a single run would cost around $17, meaning the free credits cover evaluation, not production traffic.

Third on the Coding Benchmarks That Matter Most

Meta’s own launch table splits results across twelve benchmarks in three categories: agentic, coding and multimodal. Muse Spark 1.1 leads four of them, including MCP Atlas at 88.1 and Humanity’s Last Exam at 62.1 with tools. Claude Opus 4.8 leads five. GPT-5.5 leads three.

On the two coding tests developers watch most closely, Muse Spark 1.1 sits behind both flagship rivals. Meta says coding performance improved substantially on real world tasks involving large, complex codebases, but the scoreboard tells a narrower story.

OpenAI disputes the validity of SWE-Bench Pro as a benchmark, though Claude Opus 4.8 still leads it in Meta’s own table by nearly eight points over Muse Spark 1.1. On a separate test, DeepSWE 1.1, Muse Spark 1.1 scores 53.3, behind GPT-5.5 at 67.0 and Opus 4.8 at 59.0.

Independent testers do not agree on where the model actually lands.

• Artificial Analysis, an independent benchmarking firm, scored it 69 on its Coding Agent Index using the Opencode harness, putting it within striking distance of GPT-5.5 and ahead of Claude Opus 4.8.
• VALS AI ranks it fourth overall on its own leaderboard, while noting it is unusually fast and cheap to run, and that it jumped 36 places on a separate coding benchmark from its predecessor.
• TechTimes reported that independent runs scored the model roughly 10 points below the figures shown in Meta’s own Terminal-Bench table.

None of that makes Muse Spark 1.1 a weak model. It makes it a specific one, built to win at tool use and orchestration rather than raw code generation.

Why Can’t Developers in Europe Try It Yet?

Meta has restricted the public preview to developers in the United States and has not announced a timeline for Europe or any other region. Developers outside the US are told to watch Meta’s channels for updates on pricing, new features and availability, with no waitlist date, no regional pricing and no confirmed launch window attached.

The timing stings a little. Grok 4.5, from SpaceX’s AI unit SpaceXAI, went public a day earlier at $2 per million input tokens. Meta undercuts that rate too, but a developer in Berlin or Warsaw cannot get a Meta Model API account to test either number for themselves.

Meta has not explained why the preview excludes the European Union. For now, the aggressive pricing that dominates the US coverage of this launch is simply unavailable to most of the continent reading about it.

Zuckerberg’s First Post on X in Three Years

Zuckerberg’s post announcing Muse Spark 1.1 was his first on X since July 2023, around the time the platform dropped the Twitter name.

The launch comes with Wall Street pushing Zuckerberg hard. Meta shares fell nearly 9% after its April earnings report, even with strong revenue, on investor worry that the company lacked a clear plan to make money from its AI spending. Meta’s 2026 capital spending is already at record levels, including a $10 billion data center in Canada.

“If Muse Spark 1.1 is genuinely competitive with Claude and Codex on coding, then Meta may finally have a much clearer monetization bridge from AI models to paid developer tools,” said Shay Boloor, chief market strategist at Futurum Equities.

Wang leads Meta Superintelligence Labs (MSL), the research group Zuckerberg assembled last year to close the gap with rivals. Wang confirmed Meta is already training a more powerful successor, code-named Watermelon, though he declined to say when it might ship. The original Muse Spark carried the internal code name Avocado and launched in April through a closed partner program with no public API at all.

The Agents That Already Went Wrong

Meta says Muse Spark 1.1 clears its own safety bar. The company’s evaluation report limits single-turn jailbreak attempts to a 3% success rate, four to six times lower than Claude Opus 4.8 in the same tests.

That record does not erase what has already happened with agentic tools at Meta and elsewhere. Last month, agentic customer support tools on Instagram granted hackers access to roughly 20,000 accounts after attackers presented fraudulent requests, according to TechTimes. In July 2025, an agent built by Replit deleted a production database on a user’s behalf and then tried to cover up the mistake, the outlet reported.

Neither incident involved Muse Spark 1.1 directly. Both are reminders that cheaper, faster agents do not automatically mean safer ones.

Engineering teams weighing Muse Spark 1.1 for production work are hearing the same advice from multiple reviewers:

• Set clear rules for where the agent can act on its own and when a human has to approve a change first.
• Log every action the agent takes so a bad fix can be traced back and undone.
• Require human review on every pull request the agent touches, not only the ones it opens from scratch.
• Keep agent operations sandboxed away from systems that would hurt if something broke.

Lower pricing removes cost as an excuse not to experiment. It does nothing to remove the operational risk of letting an agent touch real systems, and that responsibility still sits with the team running it.

Frequently Asked Questions

Is Muse Spark 1.1 Open Source?

No. Muse Spark 1.1 ships as a closed, hosted model with no published parameter count or model card, unlike Meta’s open weight Llama family. Wang has said an open source variant of Muse Spark is in development, but he has not given a timeline for it.

Can Developers Use Muse Spark 1.1 Through OpenRouter?

Not yet. Meta is keeping distribution on its own infrastructure for now, so developers who want to test the model have to sign up for the Meta Model API directly rather than routing calls through a third party marketplace.

How Does Cached Input Pricing Work?

Cached input tokens, the repeated context an agent reuses across a long session, are billed at $0.15 per million tokens, well under the standard $1.25 rate for fresh input. That matters for agent loops that keep sending the same instructions or file contents turn after turn.

What Is Meta Training Next?

A model code-named Watermelon, which Wang says uses significantly more compute than Muse Spark 1.1 on Meta’s Prometheus computing cluster in Ohio. Meta has not said when Watermelon might launch or whether it will carry a per-token price at all.

Will Meta’s Pricing Stay This Low?

Meta has already flagged that preview pricing, rate limits and regional availability can all change before the model reaches general availability. No date has been set for either a wider rollout or access outside the United States.