As it continues to target cybersecurity, AI lab Anthropic on Tuesday introduced two new models, demonstrating how it is improving capabilities such as reasoning and providing tools to help cyber defenders.
Claude Fable 5 is a Mythos-class model, according to Anthropic — a reference to the original Mythos, which is still in restricted release due to its reputed ability to easily find and exploit security holes in software. While the model is adept at cybersecurity tasks, Anthropic has restricted its use due to the risk of misuse as a cyber threat.
As a general-use model, Fable 5 shows impressive performance in areas such as software engineering, knowledge work, vision and scientific research. The AI lab also introduced Claude Mythos 5, which will be deployed through Project Glasswing, a vendor-led security initiative in collaboration with select companies and the U.S. government.
The new Anthropic models show how discerning enterprises need to be when deciding which models to use for their applications. While Fable 5 offers better reasoning than other frontier models, its powerful capabilities come at a tradeoff.
“[Enterprises need] to understand where Fable will make sense to bring,” said Juan Pablo Flores Cortes, a member of code review company CodeRabbit’s developer experience team. “Fable is a very powerful model for everyday tasks, but that extra power comes at a cost as well.”
Fable 5 costs $10 per million input tokens and $50 per million output tokens. In comparison, Claude Opus 4.8 costs $5 per million input tokens and $25 per million output tokens. GPT-5.5 from Anthropic rival OpenAI costs $5 per million input tokens and $30 per million output tokens. And for Google’s Gemini 3.1 Pro, the costs are between $2 and $4 per million for input tokens and between $12 and $18 per million for output tokens.
“For any business, consider the quality versus the price, and the time that you can wait for that task to be done,” Cortes added. He said it is better to use a simpler model if a business wants to perform a simple task.
CodeRabbit’s Experience
The company had early access to Fable 5, and the team found that while the model excels in long-horizon tasks (jobs that take multiple steps or a long time to complete, such as modernizing a codebase), it took twice as long on complex tasks.
“The model is taking a little bit longer because it likes to explore the current project first, understand where it’s sitting, and then start working on reasoning for a solution,” Cortes said.
In alternative, with medium-complexity software and agent tasks on which extra reasoning is helpful but not necessary, the model did quite well, he continued.
“Fable performed really well to fill the gaps in the prompts that we were providing and went even above and beyond to ensure the project was something that I would happily share in production,” he said.
A Focus on Reasoning
Fable 5 is an example of how frontier model providers are placing greater emphasis on reasoning, said Mark Beccue, an analyst at Omdia, a division of Informa TechTarget.
“Agentic is making us move into the reasoning parts,” Beccue said. He added that the rapid spread of agentic AI is why models are getting more expensive: they emphasize reasoning, which means greater reliance on inference.
“These new models, and Gemini’s in there with it, they are doing massive inference time compute, a lot of it,” Beccue said. “You must think before you respond, so that takes a lot of computing.”
He added that, compared to Google, for example, Anthropic is more focused on long-form, structured thinking and careful reasoning. This focus makes sense since Anthropic is targeting cybersecurity and code development as among the top applications for its models.
The Open Source Cost
Meanwhile, since advanced reasoning means more expensive models, it is possible that it will be harder for open source model providers to provide similarly capable models, Beccue said. Open-source AI has been touted as a less expensive and, in many cases, more flexible option for enterprise AI developers than proprietary, closed models from top generative AI vendors.
“I do not know if we can get there with open source right now in the time we need; we might need to go to proprietary models to do this. Just because it is going to take more time,” he said, adding that open source tends to be local and decentralized.
While Anthropic released Fable as a general-use model, it still included restrictions because the model could be misused in areas such as cybersecurity. This means that the model does not perform in response to or answer certain prompts Alternatively, Mythos 5 does not have the same restrictions, but access has been granted to only a few organizations.
