Anthropic Slashes AI Agent Costs by 75% While Launching Twin ‘Mythos’ Models

SAN FRANCISCO — Artificial intelligence research company Anthropic has introduced a major cost reduction for enterprise developers alongside the launch of its updated flagship artificial intelligence models, Claude Fable 5.1 and Claude Mythos 5.1. Under the new pricing, cache reads drop from $1.00 per million tokens to $0.25 per million tokens, while base pricing remains flat at $10.00 per million input tokens and $50.00 per million output tokens.

The update delivers a 75 percent reduction in prompt cache read costs for software developers using Anthropic’s Application Programming Interface (API), significantly altering the economics of running complex, multi-step AI agents and large-scale software engineering workloads. Prompt caching allows developers to store persistent contextual data—such as entire software codebases, legal libraries, or system instructions—in API memory across consecutive interactions, avoiding the compute expense of re-parsing input data on every request.

According to figures released by the San Francisco-based company, the prompt caching price drop translates to an average operational cost savings of 25 percent for standard enterprise workloads. For heavy agentic deployments, software repository analysis, and continuous coding systems that repeatedly reference long context windows, Anthropic projects operational costs will fall by up to 45 percent.

Alongside the pricing structure adjustments, Anthropic confirmed a key architectural detail regarding its flagship product strategy: Claude Fable 5.1 and Claude Mythos 5.1 are built on identical base neural network weights. In previous software generations, Anthropic referred to Fable models as belonging to the “Mythos class” without explicitly identifying them as identical core artifacts.

Founded in 2021 by former OpenAI researchers including Chief Executive Officer Dario Amodei and President Daniela Amodei, Anthropic operates as a public benefit corporation. The company built its architectural framework around “Constitutional AI,” a training method that uses a predefined set of principles to evaluate and correct AI behavior during training, alongside strict evaluation criteria detailed in its Responsible Scaling Policy.

The operational difference between the two designations lies in their deployment safeguards and accessibility profiles. Claude Fable 5.1 is equipped with hardened, automated safety guardrails designed for public commercial availability and general application deployment. Claude Mythos 5.1 retains the raw underlying capabilities with restricted access, available exclusively to pre-vetted corporate partners and vetted research organizations.

For end users, Fable 5.1 is immediately available across the Claude web interface and mobile applications. Access is integrated into the Claude Max subscription tier. Subscribers on the Claude Pro plan can access the model by purchasing separate API usage credits beyond standard subscription allowances.

The Mythos 5.1 baseline demonstrates expanded capabilities in highly specialized domain modeling, including computational biology, molecular design, and scientific simulation. Internal evaluations noted that Mythos 5.1 displays higher advanced proficiency in areas related to chemical and biological research compared to the prior Mythos 5 release.

On performance benchmarks, Claude Fable 5.1 registered measurable gains over Fable 5 across software engineering tasks, agentic scientific research protocols, multidisciplinary reasoning evaluations, and enterprise business automation pipelines. The model demonstrated specific improvements in diagnosing and resolving root-cause software bugs across complex code repositories.

Safety evaluations conducted on the system concluded that while Mythos 5.1 shows increased performance in scientific domain analysis, its capabilities in potential chemical and biological weapons synthesis do not surpass the mandatory threshold that would require elevated safety protocol lockouts beyond existing controls. Similarly, evaluations of the model’s offensive cybersecurity capabilities showed performance increases, though Anthropic stated that empirical testing yielded no evidence that its integrated safety guardrails can be bypassed or jailbroken.

Anthropic indicated that Fable 5.1 achieves performance equal to or exceeding Fable 5 even when configured to low or medium reasoning effort settings, attributing the performance gains to training modifications that prevent the model from taking shortcuts during task execution.