Meta's first frontier model lands two points behind Claude Opus 5.5, and Gemini 4 Argon opens to subscribers

Three labs moved the frontier this week, and none of them took the top of the board.

Meta released its first frontier models, Muse Wonder and Muse Wonder Full, on October 12, 2026. Artificial Analysis measures Muse Wonder Full at 56 on its Intelligence Index at maximum reasoning effort and Muse Wonder at 52; Claude Opus 5.5 sits at 58 and Claude Sonnet 5.5 ties at 56. Six weeks earlier the same index put Meta's Muse Spark 1.3 at 48. 12

The stronger of the two is not sold through the API. Muse Wonder Full runs only for Muse Code subscribers, while Muse Wonder is on the Meta Model API at $4 per million input tokens and $20 per million output. Heavy input use pushes Muse Wonder to $4.52 per index task — 2.8 times Muse Spark 1.3's $1.60, and more than twice Gemini 4 Argon's $1.99. 12

Google made Gemini 4 Argon globally available on October 13: on the API and in AI Studio, in Google Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode in Search for AI Pro and Ultra subscribers. Introductory API pricing is $2 per million input tokens and $10 per million output, doubling to $4 and $20 after November 30, 2026. Google's own testing claims 77.9% on DeepSWE v1.1 and 57% on Terminal-Bench 4.0. 3

Artificial Analysis puts Argon at 52.56 on the same index — within half a point of GPT-6 Astra at 52.67 and about five points below Claude Opus 5.5 at 57.62 — measured at the HIGH thinking setting. Subscribers default to MEDIUM, so the score most posts quote is not the configuration most users get. 4

Reflection released Beam, its first model, on October 16: open weights, 501B total and 23B active parameters, 30 on the Intelligence Index — second among US open-weights models, behind Meta's Muse Spark Open at 39 — and 48.5% on AutomationBench-AA, level with Qwen3.8 27B. 5

One number undercuts that. Beam scores 3.0% on Terminal-Bench 4.0, below Qwen3.8 27B at 5.6% and Muse Spark Open at 8.1%. On AA-Omniscience it answers 33% of questions correctly and hallucinates on 47%. 5

Three labs shipped. The top of the index stayed where it was: Claude Opus 5.5, at 58.