Meta’s latest AI coding model, Muse Spark 1.1, has turned a routine product refresh into a high-stakes bet that rock-bottom prices can win a crowded, expensive AI market.

April groundwork and AI catch‑up

Meta first introduced Muse Spark in April as its in‑house agentic coding model, part of a broader effort to catch up with OpenAI, Anthropic, and Google after a major AI-focused restructuring last year.¹ The initial version powered Meta AI experiences across apps like Instagram and WhatsApp, laying the foundation for a more ambitious enterprise push.²

July 9: Muse Spark 1.1 launches

On July 9, Meta unveiled Muse Spark 1.1, calling it a “step‑change” over the first generation, with better detection and fixing of complex bugs, stronger end‑to‑end agent workflows, and native multimodal perception across images, video, and documents.² TechCrunch described it as a multimodal model for “agentic coding” that can handle multistep reasoning, manage digital workflows, and deploy features in enterprise systems.³

The release coincided with the new Meta Model API, giving US developers public‑preview access and $20 in free credits.² CEO Mark Zuckerberg framed the model as “a strong agentic and coding model at a very low price,” available via Meta AI and the new API.Mark Zuckerberg (finkd)⁴

Performance claims and benchmark battles

Meta and its allies quickly leaned on benchmarks to argue competitiveness. The company highlighted Spark 1.1’s strength in agentic tasks that require planning and orchestration across external apps and services.³ A post amplified by Zuckerberg claimed the model became the new state of the art on MedScribe and TaxEval while being “10x cheaper and twice as fast” as a rival model, and topping Harvey’s Legal Agent Bench.Mark Zuckerberg (finkd)⁵ Another benchmark‑focused tweet noted Muse Spark 1.1 ranked #4 on the Vals Index, ahead of GPT‑5.5 and Grok 4.5.Mark Zuckerberg (finkd)⁶

Zuckerberg also told Bloomberg that Spark 1.1 outperformed Google’s Gemini in several categories, including agents and coding, arguing there is “a real ability to be able to offer frontier or very high-level intelligence at a much more affordable cost.”⁷

The price play and fears of a cost war

Behind the performance messaging is an aggressive pricing strategy. Meta’s chief AI officer said Spark 1.1 will cost $1.25 per million input tokens and $4.25 per million output tokens, undercutting Google and OpenAI’s flagship models and roughly in line with cheaper rivals like Claude Haiku and GPT‑5.6 Luna.³⁸ One AI practitioner, in a tweet Zuckerberg reposted, described Spark as “so … cheap I almost don’t believe it,” estimating it at one‑tenth the cost of competing systems like Fable and GPT‑5.5.Mark Zuckerberg (finkd)⁹

Business Insider reported that Meta is explicitly targeting what Zuckerberg called rivals’ “very extreme” pricing, with Spark 1.1 the company’s first AI model it will directly charge for — a potential new revenue stream alongside its core ads business.⁷ Investors appeared encouraged; Meta’s stock rose nearly 2% on launch day.⁷

Enterprise and developer reactions

For enterprises wrestling with AI bills, the timing is notable. Some firms, such as Coinbase, have begun capping engineers’ weekly AI spend as usage explodes.⁸ Early customer Cline, an AI coding startup, said Meta’s new price point makes it feasible to run heavy coding workloads at scale, calling that combination of capability and cost “rare.”⁸

Developer enthusiasm is also surfacing more broadly in the AI tool ecosystem. In a separate thread about building an AI‑driven news feed, one founder said they had fine‑tuned a model to write stories “better than Claude or GPT with extensive prompting,” signaling intensifying competition around specialized, cost‑efficient models — exactly the segment Spark 1.1 targets.clem 🤗 (ClementDelangue)¹⁰

What comes next

Meta is positioning Spark 1.1 as both a technical upgrade and a pricing provocation. The company says the model excels at long‑running tasks with a 1 million‑token context window, can delegate work to parallel sub‑agents, and is trained to use desktop, mobile, and browser interfaces — all while staying “very low cost.”Mark Zuckerberg (finkd)¹¹Mark Zuckerberg (finkd)¹² Zuckerberg has already promised “more to come soon,” hinting that Spark 1.1 may be only the opening move in a broader campaign to force prices down — and usage up — across the AI coding market.Mark Zuckerberg (finkd)¹³