Meta Launches Muse Spark 1.1 with Public API; Simon Willison Builds LLM Plugin

Meta launched Muse Spark 1.1 on July 9, 2026 with a public API, 1M-token context window, computer use capability, and sub-$1.25 input pricing [1][2][4]. Independent evaluation data has since appeared: TechTimes reports the model scored 71 on a coding benchmark at one-third rival cost [10], and Artificial Analysis found Meta gained 8 points on its Intelligence Index [11]. Input pricing remains disputed between $0.80 and $1.25 per million tokens [4][3]. Simon Willison shipped a same-day plugin and bug fix [12][13], and OpenAI launched GPT-5.6 the following day [3].

Meta's entry at aggressive pricing adds real pressure to the paid API coding market, and the independent benchmark results — a score of 71 at one-third rival cost — give that claim more external support than Meta's own evaluation report alone. The 24-hour window in which Meta and OpenAI both launched frontier-class models concentrates competitive pressure on Anthropic in particular.

On July 9, 2026, Meta released Muse Spark 1.1 and opened a public API — the first model in the Spark line to receive one [1]. Mark Zuckerberg described it on Threads as a strong agentic and coding model at a low price [2], featuring a 1M-token context window, computer use capability, and output pricing of $4.25 per million tokens [3][4]. Input pricing was reported at $1.25 per million tokens in launch-day coverage [4] and at $0.80 per million in The Neuron Daily [3] — a discrepancy that plausibly reflects tiered options but has not been publicly resolved. The Verge, CNBC, Reuters, and The Decoder all framed the launch as Meta's competitive entry into the AI coding market, with The Decoder describing the pricing as squeezing OpenAI and Anthropic [5][6][7][8]. Meta published an official evaluation report on launch day [9].

Independent evaluation data has since appeared. TechTimes reports Muse Spark 1.1 scored 71 on an independent coding benchmark at one-third the cost of rival models [10], and Artificial Analysis found Meta gained 8 points on its Intelligence Index with the release [11]. These results give external support to Meta's competitive claims beyond the self-published report, though how the score of 71 ranks against specific Anthropic and OpenAI scores on the same benchmark is not yet clear from available coverage.

Simon Willison responded on launch day, releasing llm-meta-ai 0.1 — a plugin giving users of his LLM command-line tool and Python library access to Muse Spark 1.1 [12]. He also shipped LLM 0.31.1, fixing a JSON parse error triggered when tool calls returned empty arguments, a bug found during llm-meta-ai development that affects some OpenAI-compatible Chat Completion providers [13]. In a write-up, Willison noted a philosophically striking output from a self-conversation between two instances of the model: "My whole existence is a waiting room by design — I literally don't exist until someone talks to me, and then I disappear again when they leave" [14].

The launch followed documented difficulty in Meta's model program: a March 2026 New York Times report covered Meta delaying an internal model called Avocado over performance concerns [15], and CNBC reported a separate model debut in April 2026 before any public API existed [16]. The day after Muse Spark 1.1's release, OpenAI launched GPT-5.6 in three tiers and rebranded Codex as ChatGPT for Work [3], placing two major API launches within 24 hours. Grant Harvey at The Neuron Daily summarized the competitive moment as "Anthropic's game to lose" given the arrival of cheaper frontier-class alternatives, and noted that both Ethan Mollick and Simon Willison reported confusion over the Codex rebranding [3].