Muse Spark 1.1: Meta Prices Its Coding Model at a Quarter of Anthropic, OpenAI

Meta opened paid API access to Muse Spark 1.1 yesterday, pricing output tokens at $4.25 per million, a rate CEO Mark Zuckerberg says lands at roughly a quarter of what Anthropic and OpenAI charge for their flagship models. The same week, an internal memo showed Meta plans to spend $145 billion on AI infrastructure this year. The pricing war is real. The open question is which models that “quarter” is actually being measured against.

Here’s my read: Zuckerberg’s quarter-price comparison isn’t measured against Anthropic’s current flagship Fable 5 or OpenAI’s GPT-5.6, it’s measured against older flagship pricing set before either company cut rates. If you’re already running coding agent workloads in production, run the same task batch through Muse Spark 1.1 and Sonnet 5, and count retries into your total token spend. Does your effective cost per task still land at a quarter? I’d genuinely like to see the numbers.

What Happened

Muse Spark 1.1 is the second model to ship under Alexandr Wang’s Meta Superintelligence Labs, following April’s original Muse Spark. The July 9 release is the first version aimed squarely at production use. It’s pitched as a multimodal reasoning model for agentic work, with a 1-million-token context window that Meta says actively compresses older information so later steps in a task still retain key details. The capability list covers tool use, computer use across applications, coding and debugging, visual-to-code generation, and parallel multi-agent orchestration for splitting a task into sub-agents that run at the same time.

On July 10, the paid Meta Model API opened in public preview for US developers: $1.25 per million input tokens, $4.25 per million output tokens, with $20 in free credits for new accounts. Zuckerberg told CNBC the pricing is “very aggressive,” roughly a quarter of what Anthropic and OpenAI charge for comparable models. The API ships in an OpenAI-compatible package format, meaning code already wired to GPT-series endpoints can switch over with a handful of configuration changes.

On benchmarks, Meta says Muse Spark 1.1 beats Google’s latest Gemini release on coding and reasoning, and beats older OpenAI and Anthropic models too. What’s missing from the published comparison table is Anthropic’s current flagships, Fable 5 and Mythos 5, and OpenAI’s GPT-5.6, which went fully public just last week. Fortune’s reporting flags that on at least one coding benchmark, Muse Spark 1.1 still trails all three of those actual flagships. Zuckerberg himself conceded at the launch that Meta’s models are “still partly lagging the competition” in places.

What the Numbers Actually Mean

Start by unpacking “a quarter of the price.” Anthropic’s Sonnet 5 currently runs at an introductory $2 input, $10 output per million tokens, rising to $3 and $15 after August 31. Muse Spark 1.1’s $4.25 output price is well under half of Sonnet 5’s rate, not a quarter. Measured against flagship Opus 4.8 at $25 per million output tokens, the quarter figure holds up cleanly. Zuckerberg’s chosen baseline is almost certainly the top-tier model, not the mid-tier model Muse Spark 1.1 is actually competing with for market share. It’s a smart framing choice, but anyone doing the savings math needs to know which tier is being compared.

The savings only materialize if completing the same task takes a comparable number of tokens. If Muse Spark 1.1’s accuracy on complex coding tasks trails Fable 5 or GPT-5.6, developers burn extra retries to land a usable result, and that extra token spend eats into the cheaper unit price fast. Fortune has already flagged at least one benchmark where Muse Spark 1.1 lags. Whether that gap translates into more tokens spent per completed task has no independent verification yet, and Meta hasn’t published retry rates or task-completion metrics, the numbers that would actually settle the cost question.

Zoom out to the infrastructure scale behind this. Meta plans to spend up to $145 billion on AI infrastructure this year. The same leaked memo describes its in-house “Iris” chip entering production in September, designed with Broadcom and manufactured by TSMC, targeting a doubling of total compute to 14 gigawatts by 2027. At current API pricing, monthly revenue in the tens of millions of dollars barely registers against a $145 billion infrastructure budget. Muse Spark 1.1’s job right now looks less like a standalone revenue engine and more like underwriting Meta AI app usage while testing how developers respond to aggressive pricing.

One more piece of context worth keeping in view. Meta’s Llama 4 benchmark results drew scrutiny last year after it emerged that the version submitted to the LMArena leaderboard wasn’t the same model shipped publicly. Given that history, Meta choosing to skip its two strongest rivals in this comparison table is a reason to discount self-reported wins like “beats Gemini” until someone else runs the numbers. Meta’s own roadmap already treats a future model codenamed “Watermelon” as the one meant to actually close the gap with competitors, which is itself an admission that Muse Spark 1.1 isn’t there yet.

Metrics Worth Watching Next

Three things should clarify this within a few months. First, whether third-party API traffic trackers like OpenRouter show developers routing production workloads to Muse Spark 1.1 rather than just burning through the free credits once. Second, whether independent benchmark groups put Muse Spark 1.1 head-to-head with Fable 5 and GPT-5.6 on the same table. If Meta keeps avoiding that comparison itself, that tells you the company already knows how it would look. Third, when Watermelon ships, and whether Meta finally puts it directly against both rivals’ flagships in a published benchmark.

Also worth tracking: yield and real deployment scale once the Iris chip enters production in September. Whether in-house silicon can actually carry the cost structure of this pricing war matters more than any single model launch.

If this was useful, subscribe to the newsletter for weekly AI PM insights and GenAI case studies.

Sources: Meta AI Blog, Fortune, TechTimes

---

Related reading:

• Claude Sonnet 5 Becomes Anthropic Default Model at Lower Cost
• GPT-5.6 Goes Fully Public: Why OpenAI’s 30-Day Review Isn’t a One-Off
• Grok 4.5 Undercuts Opus 4.8 by 4x: Who Blinks First in This Price War