Anthropic has expanded its Claude model family with the release of Claude Fable 5.1 and Mythos 5.1. The company describes the two as the same model exposed to different safeguard settings, a release strategy that allows one set of capabilities to serve both wide enterprise use and tightly controlled access in sensitive areas such as cybersecurity and life sciences.

The announcement, dated March 6, 2025, is built around three main points: a 75% cut in cache-read prices, a sharp improvement in agentic and scientific benchmark results, and the introduction of expanded watermarking and detection capabilities for AI-generated text under the European Union AI Act.

Anthropic, founded by former OpenAI researchers, has built its reputation on safety-focused AI. The company has consistently argued that large language models should be developed with robust guardrails, and its model families are designed around principles such as Constitutional AI. Fable 5.1 continues that line while aiming for practical deployment. Mythos 5.1, on the other hand, is a reminder that Anthropic's safety posture is not only about system-level alignment; it is also about access control.

Key facts

• Claude Fable 5.1 is generally available, while Mythos 5.1 is available only through two access programmes covering cybersecurity and life sciences.
• Cache read pricing falls by 75% to $0.25 per million tokens.
• Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1, up from 24.7% for Fable 5.
• Fable 5.1 scores 55.8% on Terminal-Bench 4.0.
• Claude output carries an EU AI Act watermark, and the detection API is in private preview for eligible organisations.
• The Mythos access programmes were built with the US government and are currently open only to selected US organisations.

Same model, two safety profiles

Unlike a typical dual release where one model is larger or more capable, Fable 5.1 and Mythos 5.1 share the same underlying weights and have been adapted through safeguards rather than separate training runs. Anthropic introduced Fable as generally available and made Mythos available only to a set of vetted users. Mythos 5.1 sits behind two access programmes: one covering cybersecurity and one covering the life sciences.

The biology access programme was built in cooperation with the US government. It is intended for organisations researching biological risks and developing defensive measures in life sciences. The cybersecurity track follows a similar pattern, giving vetted security teams access to the model for defensive work while controlling dangerous abuse at the deployment layer. Anthropic has long argued that advanced AI models need tiered access, and Mythos shows how the company is trying to operationalise that idea for high-risk industries. Limiting the release to partner organisations allows Anthropic to monitor misuse more directly than it could with a fully open model.

Cheaper cache reads change AI agent economics

For developers, the most important change is pricing. Cache read prices fall by 75%, from $1.00 to $0.25 per million tokens. Cache writes and other model pricing are not changed, but because applications built on large language models often reuse context, the effective cost of a production service can be significantly lower.

Anthropic estimates the reduction cuts typical workload costs by about 25%. For agentic workloads, where the model repeatedly reads long histories, tool definitions, or enterprise documents, the projected savings reach up to 45%. These figures are important in a market where enterprises are beginning to deploy autonomous agents at scale. Agents tend to carry the same instruction set, context, and memory across many steps, so their costs accumulate on the read side. By making cached reads dramatically cheaper, Anthropic is attempting to make longer-running agents more attractive; reduced cost-per-step could encourage developers to design agents with richer context and more persistent state.

The pricing decision also signals a shift in competition. Other frontier labs have lowered inference prices after their latest model releases, but routing more traffic through a prompt cache is another way to lower total costs. Prompt caching lets developers pre-load content such as enterprise guidelines, codebases, or conversation summaries and reuse them across calls. At $0.25 per million cached tokens, the price is low enough to keep a large context around for many turns, which may be crucial for solving multi-step coding, research, or data analysis tasks.

Benchmark jump in science-focused tasks

Anthropic claimed large gains on benchmarks oriented to scientific research and agents. Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1, a benchmark that evaluates agents using computer environments to solve scientific tasks. Its predecessor, Fable 5, scored only 24.7%, meaning Fable 5.1 more than doubled the earlier result.

The same model reaches 55.8% on Terminal-Bench 4.0, a broader agentic benchmark used to measure how well models complete realistic terminal-based workflows. Anthropic's own numbers should always be read with care, as model developers choose evaluation sets and methodology. Still, the jump suggests meaningful advances in agent planning, tool use, long-context reasoning, and scientific instruction following, which are important categories for real-world AI use.

These benchmark results may be especially relevant to Mythos users. In cybersecurity, agents that can probe networks, read vulnerability databases, parse code, and suggest fixes are increasingly useful. In life sciences, agents need to navigate scientific papers, plan wet-lab or computational experiments, and reason about biological data. If the model really does score 52.6% on a science benchmark, it is likely to be seen as a strong foundation for those restricted deployments.

EU watermarking in the spotlight

The release also places a large European flag in the middle of Anthropic's announcement. In July, Anthropic signed the European Union's Code of Practice on trustworthy AI and transparency of AI-generated content, joining more than 190 signatories. The code is connected to a broader movement to make generated content identifiable, especially as AI text becomes harder to distinguish from human writing.

Claude's output has always been a grey area for AI detection, but with Fable 5.1 and Mythos 5.1, the output carries a watermark. The watermark is meant to follow official deadlines under Article 50 of the EU AI Act. In practice, the provision requires providers of certain AI systems to mark synthetic content in a machine-readable format and make that content detectable. The obligation became applicable on 2 August 2026, though the company is shipping the feature as part of a model launch that predates the legal deadline.

Anthropic has been careful to describe the watermark in narrow terms. It is invisible unless someone uses the correct detection tool. It does not reveal anything about the user who generated the text. In other words, the company says it is trying to mark content without creating an identity trail. This matters under European privacy law, where metadata that can be combined with user identifiers would be subject to more restrictive rules.

The detection mechanism, however, is not yet a broadly available public tool. Anthropic previously indicated it would ship a free detection API so anyone could check a piece of text. What has been shipped is more limited: a private preview for eligible organisations, including regulators, law enforcement agencies, media outlets, fact-checkers, independent researchers, educational bodies, and EU civil society groups. Enterprises with their own compliance obligations can apply for access as well.

Anthropic says it plans to expand access to detection over time, but it has not given a date. Third-party researchers may view this limited release as a bottleneck, since the most valuable use of a watermark is verification by independent observers. The tension between regulatory compliance and controlled release is not unique to this launch; when companies make abuse-prevention and detection tools widely available, they also make it easier for actors to test for weak points.

Why Mythos remains restricted

The arrival of Mythos is also clearly designed around regulatory boundaries. The access programmes for this model were built with the US government and are open only to a set of US organisations. European cyberdefenders and life scientists are not in the programme today, two years after Claude became available in the European Union.

Anthropic says it is working to include international partners, but the immediate asymmetry is evident. The watermark in Claude is present because of European law; the verification tool is in private preview; and some of the most potentially powerful versions of the model are tied to the US government selection process. That leaves European institutions in an uncomfortable position. They may be expected to detect AI-generated content from Claude models, but they do not yet have access to the restricted variant intended for their own defensive researchers.

During the launch, the company has presented Mythos as a precautionary way to put the model in the hands of people who will apply it to defence and safety problems while controlling access to dangerous settings. Yet the official access programmes are asymmetric, and European officials are likely to press for more equal participation as the model develops.

Competitive positioning and early access

The combined release helps Anthropic address three market pressures at once. First, it gives existing enterprise users a price cut that cannot be matched by creativity in prompt compression. Second, it gives the company a credible benchmark story in agentic and science tasks, where many customers are looking for measurable improvements. Third, it lets Anthropic demonstrate compliance with Europe's AI transparency laws at a moment when every frontier model is under examination.

Enterprises building on Fable 5.1 can expect financial savings when their applications use prompt caching heavily. Those working with regulatory disclaimers or user-generated content may need to think about what watermark detection means for their workflows. At the same time, the restricted distribution of Mythos 5.1 could frustrate non-US institutions hoping to use an advanced model for defensive security or biology.

Anthropic has not announced when Mythos 5.1 might reach broader audiences. It has said access will expand over time. Until then, the launch offers a carefully balanced combination of transparency and restriction: one model with a visible price cut and watermarked output in Europe; another model with the same core abilities and extra safeguards, only for a smaller set of approved users.

Source: TNW | Anthropic News