Meta spent years giving its AI models away. On Thursday, July 9, that ended. The company released Muse Spark 1.1, a multimodal reasoning model from Meta Superintelligence Labs, and paired it with the public preview of the Meta Model API — the first time in the company's history that outside developers have been asked to pay to use a Meta model. The rate card is the story: $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits for every new account.
_(cropped).jpg){kind=link}
That number lands in a market where comparable frontier tiers sit several times higher. Mark Zuckerberg, who returned to X to post about the launch, described Muse Spark 1.1 as a strong agentic and coding model at a very low price, and said it is strongest at agent performance, tool use, and computer use. Meta's chief AI officer Alexandr Wang told CNBC the pricing is "very aggressive and attractive", and that the company wants a price that scales with heavy consumption.
The preview is US-only at launch and aimed squarely at one workload: coding agents. Meta says the model can write and debug code, drive software tools, read text, images, video and documents, and carry multi-step tasks with less human direction. It manages a 1 million-token context window that it actively compacts, remembering earlier actions and pulling back the steps that still matter.
Why It Matters
For a decade Meta's AI strategy was Llama: release open weights, let the ecosystem build on them, and capture the strategic benefit rather than the revenue. Charging for Muse Spark 1.1 is a different posture entirely. It puts Meta in the same commercial lane as OpenAI and Anthropic, competing on a rate card rather than on goodwill — and it means Meta now has a direct financial stake in how many tokens developers burn.
The timing was not subtle. Muse Spark 1.1 arrived on the same day OpenAI made its GPT-5.6 family generally available, and a day after xAI opened Grok 4.5 to developers. For the first time, several frontier-class labs shipped inside a single 48-hour window, and the differences between them showed up less in benchmark charts than in the price column. Meta chose to be the cheapest name on that list.
That choice signals something about where the industry thinks the fight is going. Raw capability at the top of the market is converging; what is not converging is the cost of running an agent that makes hundreds of tool calls per task. When an agent loops — reading a file, running a test, reading the failure, editing the code, running it again — token consumption stops being a rounding error and becomes the budget line that decides whether the product ships.
What the Pricing Actually Says
Look at what $1.25 / $4.25 implies. Frontier tiers from the largest labs have been sitting around $5 input and $25–$30 output per million tokens. Meta is asking for roughly a quarter of that on output — the side of the ledger that dominates agentic workloads, because an agent that plans, delegates and writes code generates far more than it reads.
Price cuts at this scale are rarely just marketing. They are a claim about inference economics: about how efficiently a model can be served, how much of the compute stack a company owns, and how much margin it is willing to give up to buy distribution. Meta operates its own data centers and its own silicon roadmap, and it is not trying to fund the model business out of the model business. It can afford to treat the API as an on-ramp rather than a profit center in a way that a pure-play lab, whose entire revenue line runs through that rate card, structurally cannot.
The counter-argument is that cheap tokens are only cheap if the model finishes the job. An agent that needs three attempts at $1.25 is not cheaper than one that needs a single pass at $5. The real unit of cost is the completed task, not the token, and that number will only become visible once teams run these models against their own codebases for a few months.
The Reaction
Meta lined up early partners who work in exactly the territory the model is aimed at. Replit CEO Amjad Masad praised the combination of a million-token context, multimodal support, built-in search with citations, structured output and parallel tool calling as a complete agentic foundation in a clean, OpenAI-compatible package. Cline's Saoud Rizwan pointed at the same pairing of tool use and price, calling it rare. Box's VP of AI products said the model was competitive with leading frontier models on the company's enterprise evaluation set.
The OpenAI-compatible surface matters more than it sounds. It means a team already running an agent against another provider can, in principle, change a base URL and a key and start comparing. Meta is not asking developers to adopt a new framework or rewrite their harness — it is asking them to run a bake-off, which is a much easier thing to say yes to.
The more skeptical read from engineering leaders has been about governance rather than capability. Cheaper agents mean more agents running against real repositories, and access is not the same as safety. A coding agent that edits the wrong file or misses context a human reviewer would catch does not become harmless because the tokens were inexpensive. Approval gates, scoped permissions and audit logs are the boring infrastructure that has to arrive alongside the cheap model, and it usually arrives late.
What Comes Next
The preview's limits are the obvious next questions. It is US-only, and Meta has not said when it opens more broadly. It is a preview, which means rate limits, uptime and support are still being tuned. And Meta says more capable models are already in training — the version numbering suggests the company plans to iterate quickly rather than treat 1.1 as a plateau.
Watch two things over the next quarter. First, whether the other labs answer on price. A quarter-price frontier tier from a company with Meta's balance sheet is the kind of move that forces a response, and the cleanest response is a cheaper mid-tier — something OpenAI already gestured at with the lower rungs of its GPT-5.6 lineup. Second, whether developers actually stay after the bake-off. Switching costs in this market are low on paper and high in practice: prompts get tuned to a model's quirks, evaluation suites get built around its failure modes, and inertia sets in fast.
There is also the question Meta has not answered publicly. Charging for a model changes what open weights mean for the company. Llama built enormous goodwill in the developer community precisely because it was free to run. If the best Meta models now sit behind a metered API, the open-weight story becomes a legacy commitment rather than a strategy — and the community that grew up on it will notice.
Closing Thoughts
The most interesting thing about this launch is not the model. It is that a company with 3 billion users and its own data centers decided the frontier is now cheap enough to sell at a discount, and that selling it is worth more than giving it away.
That is a statement about maturity. Frontier intelligence used to be a scarce good you rationed; Meta is treating it like a utility you meter and price to move volume. Whether that reads as commoditization or as confidence depends on how you feel about the underlying models converging — and on this particular week in July, with three labs shipping inside 48 hours and competing mostly on cost, convergence is getting hard to argue with.
한글 요약
메타가 7월 9일 새 멀티모달 추론 모델 뮤즈 스파크 1.1(Muse Spark 1.1)을 공개하면서, 동시에 메타 모델 API 공개 프리뷰를 시작했습니다. 회사 역사상 처음으로 외부 개발자에게 자사 모델 사용료를 받는 것으로, 그동안 라마(Llama) 오픈 웨이트를 무료로 배포해 온 전략에서 방향을 튼 셈입니다. 가격은 100만 입력 토큰당 1.25달러, 100만 출력 토큰당 4.25달러이며 신규 계정에는 20달러의 무료 크레딧이 제공됩니다. 현재 프리뷰는 미국 개발자 대상입니다.
핵심은 가격입니다. 경쟁 프론티어급 모델들이 대체로 입력 5달러·출력 25~30달러 선인 것과 비교하면, 출력 기준으로 약 4분의 1 수준입니다. 마크 저커버그는 X 게시글에서 "매우 낮은 가격의 강력한 에이전트·코딩 모델"이라고 소개했고, 알렉산드르 왕 최고 AI 책임자는 CNBC에 "매우 공격적이고 매력적인" 가격이며 대규모 소비량에 맞춰 확장되는 가격을 원한다고 밝혔습니다. 모델은 100만 토큰 컨텍스트 창을 스스로 관리하며, 도구 사용·컴퓨터 조작·코딩 같은 에이전트 작업에 최적화됐다고 합니다. 리플릿·클라인·박스 등 초기 파트너들은 긴 컨텍스트와 병렬 도구 호출, OpenAI 호환 인터페이스를 강점으로 꼽았습니다.
같은 주에 OpenAI가 GPT-5.6 제품군을, xAI가 그록 4.5를 잇달아 공개하면서 프론티어 모델들이 48시간 안에 몰렸고, 경쟁의 축이 성능보다 비용으로 옮겨가는 흐름이 뚜렷해졌습니다. 다만 토큰이 싸다고 작업이 싸지는 것은 아닙니다. 세 번 시도하는 저가 모델이 한 번에 끝내는 고가 모델보다 비쌀 수 있고, 실제 코드베이스에 에이전트를 붙이려면 권한 범위·승인 절차·감사 로그 같은 안전장치가 함께 필요합니다. 진짜 비용 단위는 토큰이 아니라 '완료된 작업'이며, 그 숫자는 각 팀이 몇 달 돌려본 뒤에야 드러날 겁니다.
참고: Meta AI Blog · eWeek · CNBC