Claude Fable 5.2 takes the top spot on a key AI scorecard, passing Google's Gemini

A new name at the top

Anthropic's Claude Fable 5.2 has moved ahead of Google's Gemini 4 Carbon on the Artificial Analysis Intelligence Index, as pymnts.com reported. The index is an independent scorecard, and both companies point to it when they talk about their models. Fable 5.2 scored 65, Gemini 4 Carbon scored 60, and Anthropic's own Claude Opus 5.5 scored 59.

Anthropic released the model on Monday, Nov. 9. The lead on this kind of table moves often, and pymnts.com described the frontier as having changed hands again this week. Think of a league table: the team on top today may not be there next month.

What Anthropic says is better

Beyond the score, Anthropic makes two claims about Fable 5.2. It says the model writes its output up to 30% faster than Fable 5.1. It also says its safety filters now block about 55% fewer requests that were actually legitimate, which means fewer cases where the assistant refuses a reasonable job.

Pricing works on tokens, the small chunks of text AI companies use to measure usage. Developers pay $8 for every million tokens they send to the model and $40 for every million it sends back.

On customers, Anthropic says quantitative research teams at trading and investment firms, Jump Trading among them, are using it. It also described a test at Box, where the model reviewed loan application files and caught a mismatch between salary deposits and the balance on a bank statement.

The powerful model you cannot have

Anthropic also released Claude Mythos 5.2, a sibling model restricted to vetted organizations in cybersecurity and life sciences. OpenAI took a similar route with GPT-6 Cyber, released Nov. 5 to approved customers in its Daybreak Red security program. The public cannot use it.

The reason is what the model can do. In OpenAI's testing, it found 81 serious security flaws in Firefox that were not known before, and it produced 948 working attacks against the Linux kernel, the core of the Linux operating system. OpenAI rated it Critical for cybersecurity under its Preparedness Framework, its internal system for grading risk. It is the first OpenAI model to reach that rating, and the rating is why access is gated.

OpenAI also reports that the model kept inside the limits researchers set in 99.93% of its tests. That is a company figure, not an outside check.

Planners, workers and price cuts

Meta went the other way on cost. Its new Muse Ember model, released Nov. 9, charges $0.10 per million input tokens. Meta designed it to work alongside its larger Muse Wonder: Wonder plans a task, and Ember does the repetitive steps, such as calling software tools, pulling data out of documents and reading long files. It accepts text, images and audio. Meta says it costs less per task than Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6.1 Luna.

Other releases landed in the same stretch. Thinking Machines Lab, founded by former OpenAI technology chief Mira Murati, put out Inkling 2 Small. It has 594 billion parameters, the settings a model learns in training, but uses only 17 billion on any request. The company says it will publish the model's weights under an Apache 2.0 license, which permits free commercial use, after outside safety testing. Baidu released ERNIE 5.5 Flash Base with open weights as a starting point for companies to tune on their own data.

The pymnts.com piece draws a lesson for finance firms: the question is shifting from which single model to buy toward how to send each task to the cheapest model that can handle it.

How to read a ranking

A single score blends many tests into one number. It cannot tell you which assistant writes your emails best, handles your spreadsheets, or fits your budget. A five-point gap on an index is real information, but it is not a prediction about your tasks.

The more durable change is the split in the market. Top models are priced high, the strongest specialist models are locked behind vetting, and cheaper ones are being built to do the grunt work. Knowing which job needs which tier is becoming a skill in itself.