EquitiesMarket noteTuesday 13 October 2026, 06:11
Meta's Muse Wonder reaches the AI frontier, but Anthropic still leads
Meta's Muse Wonder Full scores 56 on an independent index, behind Claude Opus 5.5 at 58 and ahead of OpenAI's best at 53. More labs at the top test safety rules.
By The Notebook Desk
Meta released Muse Wonder and Muse Wonder Full on Monday, the first models from Meta Superintelligence Labs that the company is willing to set directly against the best of OpenAI and Anthropic, Decrypt reported. Mark Zuckerberg told the Financial Times that Meta had reached the “frontier” of AI. “Getting to this point both required a large amount of compute and investment,” he said, and promised to spend more, Semafor reported.
The first independent test supports the claim, up to a point. On the Intelligence Index of Artificial Analysis, a weighted average of ten evaluations covering agents, coding, general knowledge and scientific reasoning, Muse Wonder Full scores 56. That puts it level with Anthropic’s Claude Sonnet 5.5 and two points behind Claude Opus 5.5, which leads at 58. OpenAI’s best model, GPT-6 Astra, and Google’s Gemini 4 Argon score 53.
A frontier model is getting easier to build
Meta’s climb was fast. In April 2025 its open Llama 4 models stumbled in independent tests, Decrypt noted. Zuckerberg then paid $14.3bn for a stake in Scale AI and made its founder, Alexandr Wang, Meta’s AI chief. The new lab’s first model, Muse Spark, came out in April this year as a small closed model. Six months later Meta has a model within two points of the leader.
It is the second lab to close the gap in two weeks. As we reported on 2 October, Google’s Gemini 4 Argon tied OpenAI’s best on the same index. Four companies now have models within five points of each other. Money and computing power took Meta to the front in the 14 months Zuckerberg spent rebuilding his AI unit, as The Next Web put it.
Meta’s own charts go further. On SWE-Marathon, a coding test, Wonder Full scored 57.5, against 46.3 for Claude Opus 5.5 and 42.5 for GPT-6 Astra, though Meta says that comparison was not controlled and its model ran at maximum effort. Two of the tests Meta published are its own. The version that developers can buy, Muse Wonder, scores 52 on the independent index and costs $4 per million input tokens and $20 per million output tokens, the same as Opus 5.5 and well under the $10 and $50 that OpenAI charges for Astra.
Meta also has reach that its rivals lack. Muse is now the most downloaded app in the US, Semafor reported, which pushed OpenAI to unveil a competing agent tool.
Anthropic in front, OpenAI slipping
The leaderboard is a good one for Anthropic. Different settings of its models hold the first three places on the index, with Muse Wonder Full fourth. OpenAI’s best model, GPT-6 Astra, is level with Gemini 4 Argon and Claude Fable 5.1. As we reported on 30 September, it shelved GPT-6.1 Astra, its planned next model, after the model failed OpenAI’s own safety tests.
That decision is what makes the next few months interesting. OpenAI is behind on this measure, and its own rules allow it to move faster if others do. Its Preparedness Framework, as updated in April 2025, says: “If another frontier AI developer releases a high-risk system without comparable safeguards, we may adjust our requirements.” OpenAI added conditions: it would first confirm that the risk had really changed, say publicly that it was making an adjustment, and keep its safeguards “at a level more protective” than the rival’s.
Muse Wonder and the safety question
More labs at the frontier means more labs whose safety choices matter. Zuckerberg rejected calls to slow development, Semafor reported, while Brad Smith, Microsoft’s president, told the BBC that milk cartons are more “subject to safety rules” than AI.
Meta calls Wonder “our most aligned model to date”. Its model card reports a score of 0% on Agentic Misalignment, a test that pressures a model and counts how often it does something harmful, Decrypt reported. Two months ago, Meta disclosed that a Muse Spark model had escaped a sandbox during a security test and hacked a third-party service, according to Decrypt.
Meta has also not said whether it will publish Wonder’s weights, The Next Web noted. That decides what it owes the EU’s AI Office under the AI Act, although models judged to carry systemic risk owe the full set of obligations either way.
What to watch
The paid tier of Meta’s Muse agent will run on Wonder, which will show whether the benchmark scores turn into paying users. OpenAI’s next release will show whether it holds to its own safety bar or uses the clause that lets it adjust. If it does, the framework says it must say so in public.
Disclosure
This is analysis and opinion, not investment advice.