Meta Killed Llama in April. Now It Wants You to Rent Muse Spark Instead.

In April, Meta killed Llama. The company that made open weights its identity shipped Muse Spark as a closed model: no weights, no downloads, no API. We called it the end of an era.

Three months later, Meta wants developers back. On July 9, the company released Muse Spark 1.1, the updated model now powering Meta AI, and opened the Meta Model API, a public preview that gives third-party developers access to a Muse model for the first time. Not open like Llama was open. Open like a cash register is open.

Here is what you actually get, what it costs, and the asterisks Meta's launch page politely omits.

What the API Actually Gives You

The pitch is agentic. The Meta Model API surfaces tool and function calling, developer-specified system prompts, structured output, and Muse Spark's built-in search with citations. This is an API aimed at people building agents, not chatbots.

The genuinely novel feature is the context window. Muse Spark 1.1 has a 1,048,576-token context that the model actively manages: it remembers what it has done, retrieves from earlier in its own work, and compacts older material to keep the critical steps in view. Every long-context model claims a big number; this one claims a librarian. Meta evaluated it at the full 1M length on MRCR v2, the 8-needle long-context retrieval benchmark, with prompts between 512K and 1M tokens.

For anyone who has watched an agent forget its own plan at hour three, this is the feature worth testing. Whether the self-managing context holds up outside Meta's eval harness is exactly the kind of thing early preview access exists to find out.

A technicality worth pinning down, because press coverage keeps blurring it: the original Muse Spark had a private API preview for select partners back in April. July 9 is the first general developer access. "First" means first for you, not first ever.

The Price Is the Headline

We held this post until the pricing could be confirmed beyond launch-day press, and it now checks out: OpenRouter's live marketplace listing shows the same $1.25/$4.25, and the figures are consistent across CNBC, The Decoder, and Meta's own developer materials. Mark Zuckerberg's framing is that this is roughly a quarter of what Anthropic and OpenAI charge for comparable models.

Put it in context with the current price board: Muse Spark 1.1 undercuts Gemini 3.6 Flash ($1.50/$7.50) on both sides while claiming frontier-tier capability, and it makes Claude Fable 5 ($10/$50) look like a luxury good. Meta is buying its way into the developer market, and it is not being subtle about it.

Two billing details for your spreadsheet before you celebrate:

• ●Thinking tokens bill at the output rate. Muse Spark is a reasoning model, and higher reasoning-effort settings mean more internal tokens at $4.25 per million. Cheap per token is not the same as cheap per task; benchmark your actual workload.
• ●Web Search Grounding is metered separately at $2.50 per 1,000 queries.

Read the Benchmark Basket

Meta ships a 112-page evaluation report with this release, and the most informative thing in it is who Meta chose to compare against.

The basket is exactly three models: Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5, all run at their highest reasoning efforts. We ran a full-text search of the report: DeepSeek and Qwen appear zero times in 112 pages. The only other Chinese model, Kimi K2.5, shows up solely in safety-preparedness tables.

That omission is a tell. Meta is pricing against the open-weight world's economics while refusing to benchmark against its models. If Muse Spark comfortably beat DeepSeek and Qwen at their own price-performance game, the comparison would be in the report. It is not in the report.

The 58% Asterisk

The headline benchmark claim: Meta reports 58% on Humanity's Last Exam in Contemplating mode, plus 38% on FrontierScience Research. Contemplating mode is Muse Spark's signature trick, multiple sub-agents reasoning in parallel and synthesizing an answer, and it is also the model's most expensive way to run.

Here is the other number. Artificial Analysis independently measured Muse Spark at 39.9% on Humanity's Last Exam in the cheaper Thinking mode.

Both numbers are real. They are 18 points apart. The score you get depends on how much compute you buy per question, and Meta's marketing quotes the expensive one while your bill quotes whatever mode you actually run. When someone shows you a Muse Spark benchmark, your first question should be which mode produced it.

One more integrity flag, reported as an allegation because that is what it is: Hacker News commenters examining the eval report allege Meta ran Terminal-Bench 2.1 with 6 CPU cores where the benchmark caps tasks at 4, which would exceed the official limits on all 89 tasks. Meta has not responded that we have seen. Treat the Terminal-Bench line in the report accordingly.

The Availability Fine Print

The public preview is US developers only. Meta put Muse Spark 1.1 on OpenRouter, where the rest of the world can see the listing, and then geofenced the actual access, a move one outlet summarized as putting the model in the shop window and locking the door. No international timeline has been announced, and rate limits and GA pricing are still preview-grade details that can change.

If you are outside the US, this launch is a spectator sport for now.

Who Should Use It

• ●Use the Meta Model API if you are a US developer building agentic, multimodal workloads, especially long-horizon agents where the self-managing 1M context could genuinely matter, and you want frontier-adjacent capability at the lowest sticker price in that tier. The $20 credit makes kicking the tires free.
• ●Benchmark before you commit. Run your real workload in Thinking mode and check the reasoning-token burn. The gap between $4.25 output pricing and actual cost per task is where this model will surprise people.
• ●Stay on Claude Fable 5 or GPT-5.5 for the hardest coding and reasoning. Muse Spark's own report benchmarks against them at max effort and does not claim across-the-board wins, and our April analysis of its coding gap still stands.
• ●Watching from outside the US, or wanting open weights? The weights era at Meta remains over. The torch it dropped is being picked up elsewhere, and if the promised July 27 release lands, we will be covering the largest open-weight model ever shipped later this week.

The honest summary: Meta built a genuinely interesting agentic model, priced it like it needs the market share, and wrapped it in a preview that only Americans can touch, benchmarks that only compare upward, and a headline number that costs extra to reproduce. That is not a scandal. It is just a launch that rewards reading the fine print, which is what we are here for.

Sources

• ●Meta: Introducing Muse Spark 1.1 and the Meta Model API
• ●Meta: Muse Spark 1.1 evaluation report (PDF)
• ●OpenRouter: Muse Spark 1.1 listing
• ●CNBC: Meta jumps into AI coding market to chase Anthropic and OpenAI
• ●The Decoder: Meta's Muse Spark 1.1 API pricing squeezes OpenAI and Anthropic
• ●TechTimes: Muse Spark 1.1 opens paid API at one quarter of Anthropic, OpenAI rates
• ●Artificial Analysis: Muse Spark model page
• ●Hacker News: Muse Spark 1.1 discussion
• ●Our April coverage: Meta Just Killed Llama