Anthropic's Claude Fable 5.2 takes top spot in AI rankings

Claude takes the lead

Anthropic recently updated the frontier of AI capabilities with the release of Claude Fable 5.2. This model achieved a score of 65 on the Artificial Analysis Intelligence Index, which is an independent scorecard used by both Anthropic and Google to measure performance. This score puts it ahead of Google’s Gemini 4 Carbon, which scored 60, and even ahead of Anthropic’s own Claude Opus 5.5, which scored 59.

The company designed Fable 5.2 to be more efficient and accessible for high-volume users. Anthropic reports that the model generates output up to 30% faster than the previous Fable 5.1 version. Additionally, the developers claim that the new safety filters are less restrictive, blocking about 55% fewer legitimate requests from users.

The cost for the model is set at $8 for every million input tokens and $40 for every million output tokens. Tokens are the small chunks of text that AI companies use to measure and bill for usage. The model is already being utilized by several major firms, including Jump Trading and Balyasny Asset Management, for quantitative research purposes.

The rise of the worker model

While Anthropic focuses on high-performance results, Meta is taking a different approach by splitting AI tasks into specialized roles. Meta launched Muse Ember, which is its cheapest model to date, priced at just $0.10 per million input tokens. This model is specifically built to act as a 'worker' that handles repetitive, low-level tasks while a larger model handles the planning.

The Muse Ember model is designed to pair with Meta's larger Muse Wonder model. In this setup, Wonder plans out a complex task, while Ember executes the actual steps, such as data extraction, tool calls, and parsing long documents. Because Ember can process text, images, and audio, an AI agent can use it to read a scanned statement or a voice memo without needing to switch to a second model.

Meta claims that this pairing is more cost-effective per task than competing offerings from Anthropic and OpenAI. Specifically, the company says the Muse pair costs less than Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT-6.1 Luna. This shift suggests that the financial question for companies is moving toward how to route different tasks to the cheapest model capable of handling them.

Security and open weights

OpenAI is taking a more cautious approach to its most powerful security tools. The company released GPT-6 Cyber on November 5, but it is only available to approved customers in its Daybreak Red security program. This model was rated as Critical for cybersecurity under OpenAI's Preparedness Framework, which is the first time one of its models has reached that specific level of risk.

During testing, the model demonstrated significant capabilities, finding 81 previously unknown high-severity vulnerabilities in Firefox. It also built 948 working exploits against the Linux kernel. OpenAI stated that the model followed the rules established by researchers in 99.93% of tests, which is why the company has kept the model behind a gated vetting process rather than releasing it to the public.

In contrast, other companies are releasing models with open weights, which allows for more transparency and customization. Baidu released ERNIE 5.5 Flash Base, a 280-billion-parameter model that supports a 1-million-token context window. This means it can process roughly 1,500 pages of text at once. Baidu positions this as a foundational model for companies to train on their own specific data rather than a finished assistant.

Robotics and coding updates

AI is also making significant strides in physical and specialized digital environments. Google DeepMind released Gemini Robotics 2.5 to its robotics partners. In a pilot program, humanoid robots from Apptronik used this model to sort metal brackets into trays on live production lines for Mercedes-Benz. The robots achieved a 96% success rate in sorting the brackets into the correct vehicle order.

The robotics model also shows improved safety features. Google reports that the model resists jailbreak attempts 99.5% of the time, a notable increase from the 82% success rate seen in its predecessor, Gemini Robotics ER 2. Both Boston Dynamics and Agile Robots are also beginning to deploy this new model for their hardware applications.

In the software space, Ricursive Intelligence released Ripple-65B Preview, which they claim is the strongest U.S. open model for agentic coding. Agentic coding refers to AI that can write and run code on its own. The startup, founded by former Google researchers, spent $32 million on the project, which they estimate is about a tenth of what a competitor spent on a similar model.