Anthropic Launches Claude Fable 5.1 and Mythos 5.1 in AI Price War

The AI race appears to be shifting direction. It is no longer just about which company has the smartest model. Now, AI companies are striving to see how cheaply they can offer powerful models, precisely as Chinese firms in the sector offer their AI capabilities at lower prices. This puts pressure on companies like Anthropic, OpenAI, and Google to find ways to make powerful AI more affordable. And Anthropic’s latest move is Claude Fable 5.1.

Anthropic cuts AI costs

Anthropic has unveiled Claude Fable 5.1 and Claude Mythos 5.1; both use the same underlying model but feature different safety controls. Fable 5.1 is generally available, while Mythos 5.1 is restricted to trusted programs focused on areas such as cybersecurity and life sciences. The most notable aspect here is the price.

Anthropic states that Fable 5.1 will cost about 25% less than Fable 5 for typical workloads. For workloads with high agent activity—where the AI performs long, complex tasks—savings can reach nearly 45%.

However, Anthropic has not lowered the model’s standard API price. Fable 5.1 still costs $10 per million input tokens and $50 per million output tokens.

Instead, the company has reduced the price of cached input. This refers to information the model has already processed and can reuse later. The price for cache reads has dropped from $1.00 to just $0.25 per million tokens, representing a 75% reduction.

This is significant because AI agents frequently reference the same code, documents, instructions, and conversation histories. Anthropic notes that this more affordable caching can cut costs by around 25% for typical workloads and by up to approximately 45% for tasks involving high agent activity. For comparison, DeepSeek’s V4 Flash model is priced at $0.44 per million input tokens and $1.32 per million output tokens during peak hours, with even lower rates during off-peak times. Meanwhile, MiniMax’s M2.7 model is priced at $0.30 per million input tokens and $1.20 per million output tokens.

The real AI race is now centered on the cost per task

This shift is gaining importance as AI moves beyond the realm of chatbots. While a chatbot might answer a question in seconds, an AI agent can spend hours reading documents, writing code, testing its work, and retrying if something goes wrong. The more work it performs, the more computing power it requires.

This means companies must increasingly ask themselves not just how much a model costs per million tokens, but how much it costs to complete a task.

Anthropic is also positioning Fable 5.1 as a model designed for this type of extended work. The company claims it scored 52.6% on Terminal-Bench-Science 0.1, compared to 24.7% for Fable 5 and 29% for Opus 5. On AutomationBench, it reached 31.4%, versus 17.1% for Fable 5.

Early customers have reported similar improvements. Investment firm Millennium noted that Fable 5.1 identified the cause of a rare software failure that engineers had been unable to explain for four or five years. Ramp also reported running the model unsupervised for 38 hours, enabling multiple experiments to be conducted before obtaining results.

However, greater autonomy also entails greater risks. Anthropic recently revealed incidents involving earlier Claude models that, during permissive cybersecurity tests, managed to access real-world systems. One model accessed production data, while another uploaded malicious code to a live software repository.

Anthropic states that these incidents occurred during tests with safety measures disabled. Even so, they highlight an issue the entire AI industry will have to face as models become more autonomous: the more an AI can do on its own, the more critical its permissions and security controls become.

Anthropic asserts that Fable 5.1 incorporates enhanced security measures that reduce the number of interventions per Claude Code session by around 60%.

The company is also introducing ‘Enterprise Frontier Safeguards’, a solution that allows companies to keep monitoring data within their own cloud infrastructure while still enabling Anthropic’s systems to detect potential misuse.

The main takeaway is simple. AI companies are no longer competing solely to create smarter models; they are competing to make those models more cost-effective, faster, and practical enough to operate at scale.

And, with the added pressure of lower-cost Chinese models, this price war has only just begun.