Anthropic launches Claude Fable 5.1 with lower agent costs

Introduction

Anthropic has released Claude Fable 5.1 and is also introducing Mythos 5.1 to participants in Project Glasswing. The company presents the update as more than a performance refresh: it is also an attempt to address complaints about pricing, data retention, and safeguards that sometimes block benign requests. Most evidence available at launch comes from Anthropic and early-access users, so broader independent testing will be needed before the claims can be treated as settled.

Key points

• Lower pricing is central to the release. Anthropic says Fable 5.1 is about 25 percent cheaper in typical use and can cost up to 45 percent less on complex agentic workloads. The reduction is tied in part to lower prices for cached data that has already been processed and stored. That structure could matter to coding agents and automation systems that repeatedly reuse the same context across tool calls.
• The performance pitch focuses on practical work. Every CEO Dan Shipper described it as the strongest coding model his team has used, while also calling it faster and more token-efficient. Box CEO Aaron Levie said an agent using Fable 5.1 noticed subtleties and ambiguities that Fable 5 missed in the same test. These are early impressions, not independent comparative evaluations, but they point to improvements in context handling and execution.
• Mythos emphasizes reasoning efficiency. An early observation said Mythos 5.1 at a low reasoning setting matched the benchmark performance of its predecessor at the maximum setting. Mythos remains limited to Project Glasswing participants, which makes the claim difficult for the broader developer community to verify for now.
• Safeguards are being made more targeted. Anthropic says Fable 5.1 is less likely to block basic biology questions unnecessarily. Mythos 5.1 retains the previous biology restrictions. Fable 5.1 can also be used to identify software vulnerabilities, but Anthropic says tasks such as penetration testing, exploit generation, and binary-based vulnerability scanning may still be redirected to Opus models.
• The privacy change is still pending. Enterprise Frontier Safeguards is scheduled to begin rolling out later in the fall. Anthropic says customer data will be stored on the customer’s own cloud servers rather than Anthropic’s, addressing enterprise concerns about retention and control.

Why it matters

The release reflects a broader shift in model competition. For agent products, the relevant question is not only which system has the highest benchmark score, but which one can complete a long workflow at predictable cost. Cached-context pricing is particularly important because agents often reread the same instructions, documents, or intermediate results while making multiple tool calls. If Anthropic’s savings appear in real workloads, teams may have a stronger economic case for using Fable 5.1 in coding, research, and document-processing pipelines.

There are still important qualifications. A maximum saving of 45 percent applies to complex agentic work, not necessarily to every request. Early executive endorsements cannot replace independent tests of quality, latency, reliability, and total cost. Likewise, fewer false refusals do not mean that high-risk cyber capabilities have been broadly opened. Customers will also need to wait for the details and implementation of the enterprise cloud-storage program.

Anthropic is therefore trying to improve both the economics and the day-to-day usability of its models. Whether Fable 5.1 becomes a stronger choice for agents will depend on whether the pricing advantage survives production conditions and whether its more precise safeguards can remain useful without weakening risk controls.

Source: The Verge AI